fix: allow custom converter dimensions

* fix(conversion): load .mlpackage instead of unloadable .mlmodelc

The native ct.models.MLModel runtime added in the diffusers conversion
path cannot load compiled .mlmodelc directories (no Manifest.json), so
the converter's compiled output failed at load with "A valid manifest
does not exist". Both converters now return the .mlpackage directly and
the loader lists only .mlpackage. Removes the now-dead coremlcompiler
wrappers.

* chore(release): bump version to 2.0.1

.github/workflows/publish.yml

@@ -0,0 +1,25 @@
+name: Publish to Comfy registry
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - main
+    paths:
+      - "pyproject.toml"
+
+permissions:
+  issues: write
+
+jobs:
+  publish-node:
+    name: Publish Custom Node to registry
+    runs-on: ubuntu-latest
+    if: ${{ github.repository_owner == 'aszc-dev' }}
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v4
+      - name: Publish Custom Node
+        uses: Comfy-Org/publish-node-action@v1
+        with:
+          ## Add your own personal access token to your Github Repository secrets and reference it here.
+          personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}

.github/workflows/tier0.yml

@@ -0,0 +1,27 @@
+name: Tier 0 — Unit (Linux)
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+# Deps are resolved from pyproject.toml via uv, so the toolchain pins live in
+# one place. Tier 0 must run without ComfyUI; the in-tree purity gate
+# (tests/unit/test_tier0_purity.py) enforces that the suite hasn't started
+# leaking framework imports.
+jobs:
+  unit:
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+
+      - name: uv sync
+        run: uv sync --no-install-project
+
+      - name: Run Tier 0
+        run: uv run pytest -m unit tests/ -v

.github/workflows/tier1.yml

@@ -0,0 +1,23 @@
+name: Tier 1 — Smoke (macOS self-hosted)
+
+# macOS smoke tests run on the self-hosted Apple Silicon runner instead of
+# GitHub-hosted macOS (10x minute multiplier), which exhausts the included
+# Actions minutes too quickly.
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+jobs:
+  smoke:
+    runs-on: [self-hosted, macOS, ARM64, coreml]
+    timeout-minutes: 20
+    steps:
+      - uses: actions/checkout@v4
+
+      # The self-hosted runner provides uv; no setup-uv action needed.
+      - name: uv sync
+        run: uv sync --no-install-project
+
+      - name: Run Tier 1 (synthetic micro-UNet smoke)
+        run: uv run pytest -m smoke tests/ -v

.github/workflows/tier2.yml

@@ -0,0 +1,134 @@
+name: Tier 2 — M2 / ANE (self-hosted)
+
+on:
+  pull_request:
+    # `labeled` fires when run-m2 is first added; `synchronize`/`reopened`
+    # re-run on every subsequent push while the label is present, so the
+    # result tracks the PR head instead of going stale. The `if` below keeps
+    # the run gated on the run-m2 label for all pull_request events.
+    types: [labeled, synchronize, reopened]
+  schedule:
+    # Nightly at 04:00 UTC (~05/06 in PL). Keeps the M2 path honest
+    # without burning the runner on every PR.
+    - cron: "0 4 * * *"
+  workflow_dispatch:
+
+jobs:
+  m2:
+    if: |
+      github.event_name == 'schedule' ||
+      github.event_name == 'workflow_dispatch' ||
+      (github.event_name == 'pull_request' &&
+       contains(github.event.pull_request.labels.*.name, 'run-m2'))
+    # Self-hosted Apple Silicon runner. Prerequisites: COMFY_DIR pointing at
+    # a runner-owned ComfyUI clone, plus a cached SD1.5 checkpoint.
+    runs-on: [self-hosted, macOS, ARM64, coreml]
+    timeout-minutes: 90
+    steps:
+      - uses: actions/checkout@v4
+
+      # Hybrid ComfyUI strategy:
+      #   - schedule (nightly)  -> latest origin/master + ComfyUI's own
+      #     requirements.txt (constrained). Canary for upstream API breakage.
+      #   - PR label / dispatch -> the requires-comfyui version tag + the frozen
+      #     `comfy` uv group. Reproducible merge gate, immune to overnight drift.
+      - name: Resolve ComfyUI ref + mode
+        run: |
+          if [ "$GITHUB_EVENT_NAME" = "schedule" ]; then
+            echo "COMFY_MODE=latest" >> "$GITHUB_ENV"
+            echo "COMFY_REF=master"  >> "$GITHUB_ENV"
+          else
+            # requires-comfyui is a semver constraint (e.g. ">=0.3.27"); pin the
+            # gate to the matching ComfyUI release tag (vX.Y.Z).
+            VERSION="$(sed -nE 's/^requires-comfyui *= *"[^0-9]*([0-9]+\.[0-9]+\.[0-9]+).*/\1/p' pyproject.toml)"
+            if [ -z "$VERSION" ]; then echo "could not parse requires-comfyui from pyproject.toml"; exit 1; fi
+            echo "COMFY_MODE=pinned"      >> "$GITHUB_ENV"
+            echo "COMFY_REF=v$VERSION"    >> "$GITHUB_ENV"
+          fi
+
+      - name: Set up ComfyUI checkout
+        # COMFY_DIR is exported by the self-hosted runner's .env and MUST be a
+        # runner-owned ComfyUI clone (never your dev checkout — this step does
+        # git reset --hard and rewrites custom_nodes). Cloned on first run.
+        run: |
+          set -euo pipefail
+          if [ -z "${COMFY_DIR:-}" ]; then echo "COMFY_DIR unset"; exit 1; fi
+          # Init-in-place rather than `git clone`: COMFY_DIR may already hold the
+          # cached checkpoint (models/checkpoints) or converted .mlmodelc, and
+          # `git clone` refuses a non-empty target. init + fetch + `checkout -f`
+          # populates the ComfyUI tree while leaving untracked files (the
+          # checkpoint, the cached models) untouched — so setup order is free.
+          if [ ! -d "$COMFY_DIR/.git" ]; then
+            echo "initialising ComfyUI repo in $COMFY_DIR"
+            mkdir -p "$COMFY_DIR"
+            git -C "$COMFY_DIR" init -q
+          fi
+          git -C "$COMFY_DIR" remote get-url origin >/dev/null 2>&1 \
+            || git -C "$COMFY_DIR" remote add origin https://github.com/comfyanonymous/ComfyUI.git
+          git -C "$COMFY_DIR" fetch --quiet origin
+          if [ "$COMFY_MODE" = "latest" ]; then
+            git -C "$COMFY_DIR" checkout -f -B master origin/master
+          else
+            git -C "$COMFY_DIR" checkout -f "$COMFY_REF"
+          fi
+          COMFY_SHA="$(git -C "$COMFY_DIR" rev-parse HEAD)"
+          echo "COMFY_SHA=$COMFY_SHA" >> "$GITHUB_ENV"
+          echo "Tier 2 mode=$COMFY_MODE, ComfyUI \`$COMFY_SHA\`" >> "$GITHUB_STEP_SUMMARY"
+
+          # Point ComfyUI's custom-node loader at this checkout. Refresh the
+          # symlink only; refuse to clobber a real directory (guards against a
+          # COMFY_DIR that is accidentally a dev checkout).
+          NODE_LINK="$COMFY_DIR/custom_nodes/ComfyUI-CoreMLSuite"
+          if [ -e "$NODE_LINK" ] && [ ! -L "$NODE_LINK" ]; then
+            echo "ERROR: $NODE_LINK is a real directory, not a symlink."
+            echo "COMFY_DIR must be a runner-owned ComfyUI, not your dev checkout."
+            exit 1
+          fi
+          mkdir -p "$COMFY_DIR/custom_nodes"
+          ln -sfn "$GITHUB_WORKSPACE" "$NODE_LINK"
+
+      - name: Install dependencies
+        run: |
+          set -euo pipefail
+          if [ "$COMFY_MODE" = "latest" ]; then
+            # Node deps (our coremltools-9 toolchain), then ComfyUI's own
+            # requirements for the pulled SHA, capped by the toolchain ceiling.
+            uv sync
+            uv pip install -r "$COMFY_DIR/requirements.txt" \
+              -c constraints/comfy-ceiling.txt
+          else
+            # Pinned gate: the frozen group mirrors the known-good pinned SHA.
+            uv sync --group comfy
+          fi
+
+      - name: Start ComfyUI server (background)
+        run: |
+          cd "$COMFY_DIR"
+          nohup "$GITHUB_WORKSPACE/.venv/bin/python" main.py --port 8188 --cpu-vae > /tmp/comfyui-ci.log 2>&1 &
+          # Poll the HTTP endpoint for readiness — robust to startup-banner
+          # wording / colored-log changes in a floating-latest ComfyUI.
+          for _ in $(seq 1 90); do
+            if curl -sf -o /dev/null http://127.0.0.1:8188/system_stats; then
+              echo "comfy ready (ComfyUI ${COMFY_SHA:-unknown})"; exit 0
+            fi
+            sleep 2
+          done
+          echo "comfy failed to start"; tail -100 /tmp/comfyui-ci.log; exit 1
+
+      - name: Purge cached Core ML UNets (force fresh conversion)
+        # The converter skips when a model of the same name already exists. That
+        # cache key is conversion *parameters* only, not the conversion code or
+        # toolchain — so a stale model would let a conversion regression pass.
+        # Clear it so every Tier 2 run exercises the full convert -> compile ->
+        # sample path end to end.
+        run: |
+          rm -rf "$COMFY_DIR"/models/unet/*.mlpackage "$COMFY_DIR"/models/unet/*.mlmodelc || true
+
+      - name: Run Tier 2 (m2 marker)
+        # Drives the Core ML Converter node, which converts the UNet from the
+        # checkpoint on every run (cache purged above).
+        run: uv run --no-sync pytest -m m2 tests/ -v
+
+      - name: Stop ComfyUI server
+        if: always()
+        run: pkill -f "main.py.*8188" || true

.gitignore

@@ -1,3 +1,5 @@
 playground/
-experiments/
 __pycache__/
+models/
+.venv/
+test_results/

README.md

@@ -8,8 +8,8 @@ These models are designed to leverage the Apple Neural Engine (ANE) on Apple Sil
 thereby enhancing your workflows and improving performance.
 
 If you're not sure how to obtain these models, you can download them
-[here](https://huggingface.co/coreml-community) or convert your own models using
-[coremltools](https://github.com/apple/ml-stable-diffusion).
+[here](https://huggingface.co/coreml-community) or convert your own checkpoints
+directly with the conversion nodes in this suite (see [How to use](#how-to-use)).
 
 In simple terms, think of Core ML models as a tool that can help your ComfyUI work faster and more efficiently.
 For instance, during my tests on an M2 Pro 32GB machine,
@@ -81,6 +81,29 @@ These custom nodes come with a host of features, including:
 > [!NOTE]  
 > This repository will continue to be updated with more nodes and features over time.
 
+## Conversion & Acknowledgements
+
+The Core ML conversion pipeline in this repository began as an adaptation of
+Apple's [ml-stable-diffusion](https://github.com/apple/ml-stable-diffusion),
+which pioneered running Stable Diffusion on the Apple Neural Engine. The
+implementation has since diverged and no longer depends on that package:
+
+- UNet conversion runs natively on `diffusers`' `UNet2DConditionModel`.
+- The ANE-friendly attention path (`SPLIT_EINSUM`, `SPLIT_EINSUM_V2`) is
+  reimplemented as standalone `diffusers` attention processors.
+- The toolchain tracks current ComfyUI (NumPy 2, Torch 2.7, coremltools 9,
+  Python 3.12).
+
+The goal is to keep iterating on these methods independently and to explore
+support beyond SD1.5.
+
+> [!IMPORTANT]
+> **Breaking change in 2.0.0.** The converted Core ML UNet now takes
+> `encoder_hidden_states` in the native `diffusers` layout
+> `(batch, tokens, hidden)` instead of the previous
+> `(batch, hidden, 1, tokens)`. Core ML models converted with earlier versions
+> are not compatible with 2.0.0 and must be re-converted.
+
 ## Installation
 
 ### Using ComfyUI-Manager
@@ -170,8 +193,8 @@ the node name, so if the model already exists, the node will not convert it agai
     - **ckpt_name**: The name of the checkpoint to convert. This should be the name of the checkpoint file stored in the
       `models/checkpoints` directory.
     - **model_version**: Whether the model is based on SD1.5 or SDXL.
-    - **height**: The desired height of the image generated by the model. The default is 512. Must be a multiple of 8.
-    - **width**: The desired width of the image generated by the model. The default is 512. Must be a multiple of 8.
+    - **height**: The desired height of the image generated by the model. The default is 512. Any positive multiple of 8 is accepted.
+    - **width**: The desired width of the image generated by the model. The default is 512. Any positive multiple of 8 is accepted.
     - **batch_size**: The batch size of generated images. If you're planning to generate batches of images, you can try
       increasing this value to speed up the generation process. The default is 1.
     - **attention_implementation**: The attention implementation used when converting the model. Choose SPLIT_EINSUM or
@@ -370,62 +393,59 @@ The models used in this workflow are available at the following links:
 
 ![sdxl](./assets/sdxl_conversion.png?raw=true)
 
+## Quantization (opt-in)
+
+The `Core ML Converter` and `Core ML LCM Converter` nodes accept an
+optional `quantize_nbits` dropdown that runs k-means weight palettization
+(`coremltools.optimize.coreml.palettize_weights`) on the UNet before save.
+
+Values: `none` (default — no quantization, identical to unquantized
+behavior and filenames), `8`, `6`, `4`. The number is appended to the
+.mlpackage stem as `_q<bits>` so quantized and unquantized variants
+coexist on disk and in cache.
+
+### SD1.5 1×512×512 SPLIT_EINSUM tradeoffs (M2 Pro, ANE)
+
+Measured with 20 UNet forward passes at a fixed seed for the PSNR
+comparison:
+
+| nbits | size (MB) | size vs none | fwd median (ms) | PSNR vs `none` (dB) |
+|---|---:|---:|---:|---:|
+| none | 1641 | 1.000 | 197.1 | — |
+| 8    |  822 | 0.501 | 186.6 | 53.5 |
+| 6    |  617 | 0.376 | 183.0 | 40.2 |
+| 4    |  412 | 0.251 | 179.8 | 27.5 |
+
+PSNR here is computed on the raw `noise_pred` output of a single UNet
+forward at a fixed seed, not on the final decoded image — it isolates
+the quantization-induced drift from sampler / VAE noise. Final-image
+PSNR is comfortably higher (the sampler averages over 20 steps).
+
+### Recommended settings per chip / RAM
+
+- **8 GB RAM (M1 base, M2 base):** `nbits=4`. ~4× smaller model, still
+  loads, PSNR 27 dB is visually identical at SD1.5 sizes.
+- **16 GB RAM (M1/M2/M3 Pro):** `nbits=6` is the sweet spot — ~2.7×
+  smaller, PSNR 40 dB, no perceptible quality drop.
+- **32 GB+ RAM (Max / Ultra):** `nbits=8` if you want the safety
+  margin, `none` if you want bit-identical output for golden testing.
+
+The default stays `none` so existing workflows produce byte-for-byte
+identical output.
+
 ## Limitations
 
 - Core ML models are fixed in terms of their inputs and outputs.
   This means you'll need to use latent images of the same size as the input of the model (512x512 is the default for
   SD1.5).
-  However, you can convert the model to a different input size using tools available
-  in the [apple/ml-stable-diffusion](https://github.com/apple/ml-stable-diffusion) repository.
+  However, you can re-convert the model to a different input size using the
+  conversion nodes in this suite (set the desired width and height).
 - SD2.1 models are not supported.
 
 [^1]:
 Unless [EnumeratedShapes](https://apple.github.io/coremltools/docs-guides/source/flexible-inputs.html#select-from-predetermined-shapes)
 is used during conversion. Needs more testing.
 
-## FAQ
-
-### Hardware and Performance
-
-#### What's the difference between MPS, GPU, and ANE?
-- **MPS (Metal Performance Shaders)**: Apple's framework for GPU acceleration. It's what PyTorch uses by default on Apple Silicon.
-- **GPU**: The graphics processing unit on your Apple Silicon chip.
-- **ANE (Apple Neural Engine)**: A specialized hardware accelerator for machine learning tasks.
-
-#### Which compute unit should I choose?
-- **CPU_AND_ANE**: Best for models converted with `--attention-implementation SPLIT_EINSUM`. This is the default and recommended option for most users.
-- **CPU_AND_GPU**: Best for models converted with `--attention-implementation ORIGINAL`. Use this if you experience issues with ANE.
-- **CPU_ONLY**: Use this as a fallback if you experience issues with both ANE and GPU.
-
-#### Do I need `PYTORCH_ENABLE_MPS_FALLBACK=1`?
-While our Core ML nodes don't use this environment variable directly, it may still be relevant for other parts of ComfyUI that use PyTorch with MPS backend. The setting of this variable is a user preference and depends on your specific needs and workflow requirements.
-
-### Model Conversion and Compatibility
-
-#### Is there a performance penalty when using the Core ML Adapter?
-Yes, there might be a slight performance penalty compared to using directly converted models. However, the adapter provides more flexibility and compatibility with standard ComfyUI nodes.
-
-#### Does the Core ML Adapter support SDXL?
-Currently, SDXL support in the Core ML Adapter is limited. While it may work with some models, it's not officially supported and may cause issues.
-
-#### Are `mlmodelc` and `mlpackage` formats safe?
-Yes, both formats are safe to use. However, we recommend:
-1. Always downloading original `.safetensors` files from trusted sources
-2. Converting them yourself using our tools
-3. Using the converted `.mlmodelc` files for better performance
-
-#### Do Core ML models produce identical results to their safetensors counterparts?
-While the results should be very similar, there might be slight differences due to:
-- Different numerical precision
-- Hardware-specific optimizations
-- Different attention implementations
-
-#### Should I convert models every time I queue a generation?
-No! The conversion only happens once when you first use the converter node. After that, you should use the `CoreMLUnetLoader` to load the already converted model.
-
-#### Will SDXL ever be supported on ANE?
-Currently, there are technical limitations preventing SDXL from running efficiently on ANE. We recommend using `CPU_AND_GPU` or `CPU_ONLY` for SDXL models.
-
 ## Support
 
 I'm here to help! If you have any questions or suggestions, don't hesitate to open an issue and I'll do my best

conftest.py

@@ -0,0 +1,4 @@
+"""Top-level conftest: prevent pytest from importing the repo-root
+__init__.py (the ComfyUI custom-node entry point pulls in comfy + nodes,
+which breaks the Tier-0 'no-framework' promise)."""
+collect_ignore = ["__init__.py"]

constraints/comfy-ceiling.txt

@@ -0,0 +1,18 @@
+# Toolchain ceiling for installing a floating-latest ComfyUI's requirements.txt
+# in the Tier 2 nightly canary (.github/workflows/tier2.yml, latest mode).
+#
+# ComfyUI's requirements.txt requests bare `torch`/`torchvision`/`torchaudio`
+# and `numpy>=1.25.0`, which would float past the versions coremltools 9 /
+# apple-ml-stable-diffusion have been validated against.
+# These constraints cap the resolution so the canary keeps testing the same
+# toolchain the suite actually ships.
+#
+# If upstream ComfyUI ever hard-requires something beyond these bounds, the
+# install FAILS — and that failure is the signal we want: it means the host
+# outgrew the pinned toolchain and coremltools / ml-stable-diffusion need a
+# deliberate bump, not a silent float.
+torch>=2.7,<2.8
+torchvision>=0.22,<0.23
+torchaudio>=2.7,<2.8
+numpy>=1.25,<2
+coremltools>=9,<10

coreml_suite/attention.py

@@ -0,0 +1,5 @@
+ATTENTION_IMPLEMENTATIONS = (
+    "SPLIT_EINSUM",
+    "SPLIT_EINSUM_V2",
+    "ORIGINAL",
+)

coreml_suite/config.py

@@ -1,17 +1,10 @@
-from enum import Enum
-
 import torch
 
 from comfy import supported_models_base
 from comfy import latent_formats
 from comfy.model_detection import convert_config
 
-
-class ModelVersion(Enum):
-    SD15 = "sd15"
-    SDXL = "sdxl"
-    SDXL_REFINER = "sdxl_refiner"
-    LCM = "lcm"
+from coreml_suite.model_version import ModelVersion
 
 
 config_map = {

coreml_suite/controlnet.py

@@ -1,62 +1,14 @@
-from itertools import chain
-from math import ceil
+"""Compatibility shim — re-exports from coreml_suite.core.controlnet."""
+from coreml_suite.core.controlnet import (
+    chunk_control,
+    expand_inputs,
+    extract_residual_kwargs,
+    no_control,
+)
 
-import numpy as np
-import torch
-
-from coreml_suite.latents import chunk_batch
-
-
-def expand_inputs(inputs):
-    expanded = inputs.copy()
-    for k, v in inputs.items():
-        if isinstance(v, np.ndarray):
-            expanded[k] = np.concatenate([v] * 2) if v.shape[0] == 1 else v
-        elif isinstance(v, torch.Tensor):
-            expanded[k] = torch.cat([v] * 2) if v.shape[0] == 1 else v
-        elif isinstance(v, list):
-            expanded[k] = v * 2 if len(v) == 1 else v
-        elif isinstance(v, dict):
-            expand_inputs(v)
-    return expanded
-
-
-def extract_residual_kwargs(expected_inputs, control):
-    if "additional_residual_0" not in expected_inputs.keys():
-        return {}
-    if control is None:
-        return no_control(expected_inputs)
-
-    residual_kwargs = {
-        "additional_residual_{}".format(i): r.cpu().numpy().astype(np.float16)
-        for i, r in enumerate(chain(control["output"], control["middle"]))
-    }
-    return residual_kwargs
-
-
-def no_control(expected_inputs):
-    shapes_dict = {
-        k: v["shape"] for k, v in expected_inputs.items() if k.startswith("additional")
-    }
-    residual_kwargs = {
-        k: torch.zeros(*shape).cpu().numpy().astype(dtype=np.float16)
-        for k, shape in shapes_dict.items()
-    }
-    return residual_kwargs
-
-
-def chunk_control(cn, target_size):
-    if cn is None:
-        return [None] * target_size
-
-    num_chunks = ceil(cn["output"][0].shape[0] / target_size)
-
-    out = [{"output": [], "middle": []} for _ in range(num_chunks)]
-
-    for k, v in cn.items():
-        for i, x in enumerate(v):
-            chunks = chunk_batch(x, (target_size, *x.shape[1:]))
-            for j, chunk in enumerate(chunks):
-                out[j][k].append(chunk)
-
-    return out
+__all__ = [
+    "chunk_control",
+    "expand_inputs",
+    "extract_residual_kwargs",
+    "no_control",
+]

coreml_suite/conversion/__init__.py

@@ -0,0 +1,9 @@
+"""Core ML conversion helpers.
+
+The conversion approach originates from Apple's ml-stable-diffusion
+(https://github.com/apple/ml-stable-diffusion). This implementation has since
+diverged: it runs natively on diffusers' UNet2DConditionModel with its own
+SPLIT_EINSUM / SPLIT_EINSUM_V2 attention processors and no longer depends on
+that package. The intent is to keep iterating on these methods independently
+while tracking current tooling.
+"""

coreml_suite/conversion/attention.py

@@ -0,0 +1,239 @@
+import logging
+
+import torch
+
+logger = logging.getLogger(__name__)
+
+CHUNK_SIZE = 512
+
+
+def apply_attention_implementation(unet, attention_implementation):
+    if attention_implementation == "ORIGINAL":
+        return unet
+
+    if attention_implementation == "SPLIT_EINSUM":
+        unet.set_attn_processor(SplitEinsumAttnProcessor())
+        return unet
+
+    if attention_implementation == "SPLIT_EINSUM_V2":
+        unet.set_attn_processor(SplitEinsumV2AttnProcessor())
+        return unet
+
+    raise ValueError(f"Unsupported attention implementation: {attention_implementation}")
+
+
+class SplitEinsumAttnProcessor:
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        *args,
+        **kwargs,
+    ):
+        return _attention_forward(
+            attn,
+            hidden_states,
+            encoder_hidden_states,
+            attention_mask,
+            temb,
+            split_einsum,
+        )
+
+
+class SplitEinsumV2AttnProcessor:
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        *args,
+        **kwargs,
+    ):
+        return _attention_forward(
+            attn,
+            hidden_states,
+            encoder_hidden_states,
+            attention_mask,
+            temb,
+            split_einsum_v2,
+        )
+
+
+def _attention_forward(
+    attn,
+    hidden_states,
+    encoder_hidden_states,
+    attention_mask,
+    temb,
+    attention_fn,
+):
+    residual = hidden_states
+
+    if attn.spatial_norm is not None:
+        hidden_states = attn.spatial_norm(hidden_states, temb)
+
+    input_ndim = hidden_states.ndim
+    if input_ndim == 4:
+        batch_size, channel, height, width = hidden_states.shape
+        hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+    else:
+        batch_size, _, channel = hidden_states.shape
+        height = None
+        width = None
+
+    batch_size, key_sequence_length, _ = (
+        hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+    )
+
+    if attention_mask is not None:
+        attention_mask = attn.prepare_attention_mask(
+            attention_mask,
+            key_sequence_length,
+            batch_size,
+        )
+        attention_mask = _prepare_split_einsum_mask(
+            attention_mask,
+            batch_size,
+            attn.heads,
+            key_sequence_length,
+        )
+
+    if attn.group_norm is not None:
+        hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+    query = attn.to_q(hidden_states)
+
+    if encoder_hidden_states is None:
+        encoder_hidden_states = hidden_states
+    elif attn.norm_cross:
+        encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+    key = attn.to_k(encoder_hidden_states)
+    value = attn.to_v(encoder_hidden_states)
+
+    batch_size = query.shape[0]
+    dim_head = attn.inner_kv_dim // attn.heads
+
+    query = _linear_projection_to_bchw(query)
+    key = _linear_projection_to_bchw(key)
+    value = _linear_projection_to_bchw(value)
+
+    hidden_states = attention_fn(
+        query,
+        key,
+        value,
+        attention_mask,
+        attn.heads,
+        dim_head,
+    )
+    hidden_states = hidden_states.squeeze(2).transpose(1, 2)
+    hidden_states = hidden_states.reshape(batch_size, -1, attn.inner_dim)
+
+    hidden_states = attn.to_out[0](hidden_states)
+    hidden_states = attn.to_out[1](hidden_states)
+
+    if input_ndim == 4:
+        hidden_states = hidden_states.transpose(-1, -2).reshape(
+            batch_size,
+            channel,
+            height,
+            width,
+        )
+
+    if attn.residual_connection:
+        hidden_states = hidden_states + residual
+
+    hidden_states = hidden_states / attn.rescale_output_factor
+    return hidden_states
+
+
+def split_einsum(q, k, v, mask, heads, dim_head):
+    q_heads = _split_heads(q, heads, dim_head)
+    k = k.transpose(1, 3)
+    k_heads = [
+        k[:, :, :, head_idx * dim_head : (head_idx + 1) * dim_head]
+        for head_idx in range(heads)
+    ]
+    v_heads = _split_heads(v, heads, dim_head)
+
+    weights = [
+        torch.einsum("bchq,bkhc->bkhq", query, key) * (dim_head**-0.5)
+        for query, key in zip(q_heads, k_heads)
+    ]
+    if mask is not None:
+        weights = [weight + mask for weight in weights]
+
+    weights = [weight.softmax(dim=1) for weight in weights]
+    outputs = [
+        torch.einsum("bkhq,bchk->bchq", weight, value)
+        for weight, value in zip(weights, v_heads)
+    ]
+    return torch.cat(outputs, dim=1)
+
+
+def split_einsum_v2(q, k, v, mask, heads, dim_head):
+    query_length = q.size(3)
+    num_chunks = query_length // CHUNK_SIZE
+    if num_chunks == 0:
+        logger.info(
+            "SPLIT_EINSUM_V2 query sequence is shorter than %s; using SPLIT_EINSUM.",
+            CHUNK_SIZE,
+        )
+        return split_einsum(q, k, v, mask, heads, dim_head)
+
+    q_heads = _split_heads(q, heads, dim_head)
+    q_chunks = [
+        [
+            head[..., chunk_idx * CHUNK_SIZE : (chunk_idx + 1) * CHUNK_SIZE]
+            for chunk_idx in range(num_chunks)
+        ]
+        for head in q_heads
+    ]
+
+    k = k.transpose(1, 3)
+    k_heads = [
+        k[:, :, :, head_idx * dim_head : (head_idx + 1) * dim_head]
+        for head_idx in range(heads)
+    ]
+    v_heads = _split_heads(v, heads, dim_head)
+
+    head_outputs = []
+    for query_chunks, key, value in zip(q_chunks, k_heads, v_heads):
+        chunk_outputs = []
+        for query_chunk in query_chunks:
+            weights = torch.einsum("bchq,bkhc->bkhq", query_chunk, key)
+            weights = weights * (dim_head**-0.5)
+            if mask is not None:
+                weights = weights + mask
+            weights = weights.softmax(dim=1)
+            chunk_outputs.append(torch.einsum("bkhq,bchk->bchq", weights, value))
+        head_outputs.append(torch.cat(chunk_outputs, dim=3))
+
+    return torch.cat(head_outputs, dim=1)
+
+
+def _split_heads(x, heads, dim_head):
+    return [
+        x[:, head_idx * dim_head : (head_idx + 1) * dim_head, :, :]
+        for head_idx in range(heads)
+    ]
+
+
+def _linear_projection_to_bchw(x):
+    return x.transpose(1, 2).unsqueeze(2)
+
+
+def _prepare_split_einsum_mask(mask, batch_size, heads, key_sequence_length):
+    if mask.ndim == 2:
+        mask = mask[:, None, :]
+    if mask.shape[0] == batch_size * heads:
+        mask = mask.reshape(batch_size, heads, -1, key_sequence_length)
+        mask = mask[:, 0]
+    if mask.ndim == 3:
+        mask = mask[:, :, None, None]
+    return mask

coreml_suite/conversion/shapes.py

@@ -0,0 +1,20 @@
+def conv2d_output_shape(height, width, conv):
+    """Return the spatial output shape for a torch.nn.Conv2d-like module."""
+    kernel_h, kernel_w = _pair(conv.kernel_size)
+    stride_h, stride_w = _pair(conv.stride)
+    pad_h, pad_w = _pair(conv.padding)
+    dilation_h, dilation_w = _pair(conv.dilation)
+
+    out_h = _conv_output_dim(height, kernel_h, stride_h, pad_h, dilation_h)
+    out_w = _conv_output_dim(width, kernel_w, stride_w, pad_w, dilation_w)
+    return out_h, out_w
+
+
+def _conv_output_dim(size, kernel, stride, padding, dilation):
+    return ((size + (2 * padding) - (dilation * (kernel - 1)) - 1) // stride) + 1
+
+
+def _pair(value):
+    if isinstance(value, tuple):
+        return value
+    return value, value

coreml_suite/conversion/trace.py

@@ -0,0 +1,61 @@
+from types import MethodType
+
+from diffusers.models.transformers.transformer_2d import Transformer2DModel
+
+
+def prepare_unet_for_coreml_trace(unet):
+    for module in unet.modules():
+        if isinstance(module, Transformer2DModel):
+            module._operate_on_continuous_inputs = MethodType(
+                _operate_on_continuous_inputs,
+                module,
+            )
+            module._get_output_for_continuous_inputs = MethodType(
+                _get_output_for_continuous_inputs,
+                module,
+            )
+    return unet
+
+
+def _operate_on_continuous_inputs(self, hidden_states):
+    hidden_states = self.norm(hidden_states)
+
+    if not self.use_linear_projection:
+        hidden_states = self.proj_in(hidden_states)
+        inner_dim = self.inner_dim
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)
+    else:
+        inner_dim = hidden_states.shape[1]
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)
+        hidden_states = self.proj_in(hidden_states)
+
+    return hidden_states, inner_dim
+
+
+def _get_output_for_continuous_inputs(
+    self,
+    hidden_states,
+    residual,
+    batch_size,
+    height,
+    width,
+    inner_dim,
+):
+    if not self.use_linear_projection:
+        hidden_states = hidden_states.transpose(1, 2).reshape(
+            batch_size,
+            inner_dim,
+            height,
+            width,
+        )
+        hidden_states = self.proj_out(hidden_states)
+    else:
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = hidden_states.transpose(1, 2).reshape(
+            batch_size,
+            inner_dim,
+            height,
+            width,
+        )
+
+    return hidden_states + residual

coreml_suite/conversion/unet.py

@@ -0,0 +1,54 @@
+import torch
+
+
+class CoreMLUNetWrapper(torch.nn.Module):
+    """Adapt diffusers UNet inputs to CoreMLSuite's stable Core ML contract."""
+
+    def __init__(self, unet, model_version):
+        super().__init__()
+        self.unet = unet
+        self.model_version = model_version
+
+    def forward(self, sample, timestep, encoder_hidden_states, *extra_inputs):
+        input_index = 0
+        timestep_cond = None
+        if self._is_lcm:
+            timestep_cond = extra_inputs[input_index]
+            input_index += 1
+
+        added_cond_kwargs = None
+        if self._is_sdxl:
+            time_ids = extra_inputs[input_index]
+            text_embeds = extra_inputs[input_index + 1]
+            input_index += 2
+            added_cond_kwargs = {
+                "time_ids": time_ids,
+                "text_embeds": text_embeds,
+            }
+
+        additional_residuals = extra_inputs[input_index:]
+        down_residuals = None
+        mid_residual = None
+        if additional_residuals:
+            down_residuals = tuple(additional_residuals[:-1])
+            mid_residual = additional_residuals[-1]
+
+        outputs = self.unet(
+            sample,
+            timestep,
+            encoder_hidden_states=encoder_hidden_states,
+            timestep_cond=timestep_cond,
+            added_cond_kwargs=added_cond_kwargs,
+            down_block_additional_residuals=down_residuals,
+            mid_block_additional_residual=mid_residual,
+            return_dict=False,
+        )
+        return outputs[0]
+
+    @property
+    def _is_lcm(self):
+        return self.model_version.name == "LCM"
+
+    @property
+    def _is_sdxl(self):
+        return self.model_version.name in {"SDXL", "SDXL_REFINER"}

coreml_suite/converter.py

@@ -1,72 +1,38 @@
 import gc
 import os
-import shutil
 import time
-from typing import Union
 
 import coremltools as ct
 import numpy as np
-import python_coreml_stable_diffusion.unet
 import torch
-from diffusers import (
-    StableDiffusionPipeline,
-    LatentConsistencyModelPipeline,
-    StableDiffusionXLPipeline,
-)
-from python_coreml_stable_diffusion.unet import (
-    UNet2DConditionModel,
-    UNet2DConditionModelXL,
-    AttentionImplementations,
-)
+from diffusers import UNet2DConditionModel
 
-from coreml_suite.config import ModelVersion
-from coreml_suite.lcm.unet import UNet2DConditionModelLCM
+from coreml_suite.attention import ATTENTION_IMPLEMENTATIONS
+from coreml_suite.conversion.attention import apply_attention_implementation
+from coreml_suite.conversion.shapes import conv2d_output_shape
+from coreml_suite.conversion.trace import prepare_unet_for_coreml_trace
+from coreml_suite.conversion.unet import CoreMLUNetWrapper
 from coreml_suite.logger import logger
-from folder_paths import get_folder_paths
+from coreml_suite.model_version import ModelVersion
+
+DEFAULT_TRACE_TIMESTEP = 999.0
+TEXT_TOKEN_SEQUENCE_LENGTH = 77
 
 
-class StableDiffusionLCMPipeline(LatentConsistencyModelPipeline):
-    pass
-
-
-MODEL_TYPE_TO_UNET_CLS = {
-    ModelVersion.SD15: UNet2DConditionModel,
-    ModelVersion.SDXL: UNet2DConditionModelXL,
-    ModelVersion.LCM: UNet2DConditionModelLCM,
-}
-
-MODEL_TYPE_TO_PIPE_CLS = {
-    ModelVersion.SD15: StableDiffusionPipeline,
-    ModelVersion.SDXL: StableDiffusionXLPipeline,
-    ModelVersion.LCM: StableDiffusionLCMPipeline,
-}
-
-
-def get_unet(model_type: ModelVersion, ref_pipe):
-    ref_unet = ref_pipe.unet
-
-    unet_cls = MODEL_TYPE_TO_UNET_CLS[model_type]
-    cml_unet = unet_cls.from_config(ref_unet.config).eval()
-    cml_unet.load_state_dict(ref_unet.state_dict(), strict=False)
-
-    return cml_unet
-
-
-def get_encoder_hidden_states_shape(ref_pipe, batch_size):
-    text_encoder = (
-        ref_pipe.text_encoder_2
-        if hasattr(ref_pipe, "text_encoder_2")
-        else ref_pipe.text_encoder
+def get_unet(model_version: ModelVersion, ref_unet, attention_implementation):
+    ref_unet = prepare_unet_for_coreml_trace(ref_unet)
+    unet = apply_attention_implementation(
+        ref_unet.eval(),
+        attention_implementation,
     )
+    return CoreMLUNetWrapper(unet, model_version)
 
-    text_token_sequence_length = text_encoder.config.max_position_embeddings
-    hidden_size = (text_encoder.config.hidden_size,)
 
+def get_encoder_hidden_states_shape(ref_unet, batch_size):
     encoder_hidden_states_shape = (
         batch_size,
-        ref_pipe.unet.config.cross_attention_dim or hidden_size,
-        1,
-        text_token_sequence_length,
+        TEXT_TOKEN_SEQUENCE_LENGTH,
+        ref_unet.config.cross_attention_dim,
     )
 
     return encoder_hidden_states_shape
@@ -122,38 +88,21 @@ def convert_to_coreml(
 
 
 def get_out_path(submodule_name, model_name):
+    from folder_paths import get_folder_paths
+
     fname = f"{model_name}_{submodule_name}.mlpackage"
     unet_path = get_folder_paths(submodule_name)[0]
     out_path = os.path.join(unet_path, fname)
     return out_path
 
 
-def compile_coreml_model(source_model_path, output_dir, final_name):
-    """Compiles Core ML models using the coremlcompiler utility from Xcode toolchain"""
-    target_path = os.path.join(output_dir, f"{final_name}.mlmodelc")
-    if os.path.exists(target_path):
-        logger.warning(f"Found existing compiled model at {target_path}! Skipping..")
-        return target_path
-
-    logger.info(f"Compiling {source_model_path}")
-    source_model_name = os.path.basename(os.path.splitext(source_model_path)[0])
-
-    os.system(f"xcrun coremlcompiler compile {source_model_path} {output_dir}")
-    compiled_output = os.path.join(output_dir, f"{source_model_name}.mlmodelc")
-    shutil.move(compiled_output, target_path)
-
-    return target_path
-
-
-def get_sample_input(batch_size, encoder_hidden_states_shape, sample_shape, scheduler):
+def get_sample_input(batch_size, encoder_hidden_states_shape, sample_shape):
     sample_unet_inputs = dict(
         [
             ("sample", torch.rand(*sample_shape)),
             (
                 "timestep",
-                torch.tensor([scheduler.timesteps[0].item()] * batch_size).to(
-                    torch.float32
-                ),
+                torch.tensor([DEFAULT_TRACE_TIMESTEP] * batch_size).to(torch.float32),
             ),
             ("encoder_hidden_states", torch.rand(*encoder_hidden_states_shape)),
         ]
@@ -166,7 +115,7 @@ def lcm_inputs(sample_unet_inputs):
     return {"timestep_cond": torch.randn(batch_size, 256).to(torch.float32)}
 
 
-def sdxl_inputs(sample_unet_inputs, ref_pipe):
+def sdxl_inputs(sample_unet_inputs, ref_unet, model_version):
     sample_shape = sample_unet_inputs["sample"].shape
     batch_size = sample_shape[0]
     h = sample_shape[2] * 8
@@ -174,10 +123,7 @@ def sdxl_inputs(sample_unet_inputs, ref_pipe):
     original_size = (h, w)
     crops_coords_top_left = (0, 0)
 
-    is_refiner = (
-        hasattr(ref_pipe.config, "requires_aesthetics_score")
-        and ref_pipe.config.requires_aesthetics_score
-    )
+    is_refiner = model_version == ModelVersion.SDXL_REFINER
 
     if is_refiner:
         aesthetic_score = (6.0,)
@@ -187,7 +133,7 @@ def sdxl_inputs(sample_unet_inputs, ref_pipe):
         time_ids_list = list(original_size + crops_coords_top_left + target_size)
 
     time_ids = torch.tensor(time_ids_list).repeat(batch_size, 1).to(torch.int64)
-    text_embeds_shape = (batch_size, ref_pipe.text_encoder_2.config.hidden_size)
+    text_embeds_shape = (batch_size, get_sdxl_text_embeds_dim(ref_unet, len(time_ids_list)))
 
     return {
         "time_ids": time_ids,
@@ -195,21 +141,25 @@ def sdxl_inputs(sample_unet_inputs, ref_pipe):
     }
 
 
+def get_sdxl_text_embeds_dim(ref_unet, time_ids_dim):
+    projection_dim = ref_unet.config.projection_class_embeddings_input_dim
+    time_embed_dim = ref_unet.config.addition_time_embed_dim
+    return projection_dim - (time_ids_dim * time_embed_dim)
+
+
 def get_inputs_spec(inputs):
     inputs_spec = {k: (v.shape, v.dtype) for k, v in inputs.items()}
     return inputs_spec
 
 
 def add_cnet_support(sample_shape, reference_unet):
-    from python_coreml_stable_diffusion.unet import calculate_conv2d_output_shape
-
     additional_residuals_shapes = []
 
     batch_size = sample_shape[0]
     h, w = sample_shape[2:]
 
     # conv_in
-    out_h, out_w = calculate_conv2d_output_shape(
+    out_h, out_w = conv2d_output_shape(
         h,
         w,
         reference_unet.conv_in,
@@ -226,9 +176,7 @@ def add_cnet_support(sample_shape, reference_unet):
         ]
         if hasattr(down_block, "downsamplers") and down_block.downsamplers is not None:
             for downsampler in down_block.downsamplers:
-                out_h, out_w = calculate_conv2d_output_shape(
-                    out_h, out_w, downsampler.conv
-                )
+                out_h, out_w = conv2d_output_shape(out_h, out_w, downsampler.conv)
             additional_residuals_shapes.append(
                 (
                     batch_size,
@@ -252,15 +200,16 @@ def add_cnet_support(sample_shape, reference_unet):
 
 
 def convert_unet(
-    ref_pipe,
+    ref_unet,
     model_version: ModelVersion,
     unet_out_path: str,
     batch_size: int = 1,
     sample_size: tuple[int, int] = (64, 64),
     controlnet_support: bool = False,
+    attention_implementation: str = ATTENTION_IMPLEMENTATIONS[0],
+    quantize_nbits: str = "none",
 ):
-    coreml_unet = get_unet(model_version, ref_pipe)
-    ref_unet = ref_pipe.unet
+    coreml_unet = get_unet(model_version, ref_unet, attention_implementation)
 
     sample_shape = (
         batch_size,  # B
@@ -269,20 +218,17 @@ def convert_unet(
         sample_size[1],  # W
     )
 
-    encoder_hidden_states_shape = get_encoder_hidden_states_shape(ref_pipe, batch_size)
-
-    scheduler = ref_pipe.scheduler
-    scheduler.set_timesteps(50)
+    encoder_hidden_states_shape = get_encoder_hidden_states_shape(ref_unet, batch_size)
 
     sample_inputs = get_sample_input(
-        batch_size, encoder_hidden_states_shape, sample_shape, scheduler
+        batch_size, encoder_hidden_states_shape, sample_shape
     )
 
     if model_version == ModelVersion.LCM:
         sample_inputs |= lcm_inputs(sample_inputs)
 
-    if model_version == ModelVersion.SDXL:
-        sample_inputs |= sdxl_inputs(sample_inputs, ref_pipe)
+    if model_version in {ModelVersion.SDXL, ModelVersion.SDXL_REFINER}:
+        sample_inputs |= sdxl_inputs(sample_inputs, ref_unet, model_version)
 
     if controlnet_support:
         sample_inputs |= add_cnet_support(sample_shape, ref_unet)
@@ -305,6 +251,24 @@ def convert_unet(
     del traced_unet
     gc.collect()
 
+    if quantize_nbits != "none":
+        # Opt-in k-means weight palettization. The default path
+        # (quantize_nbits="none") leaves the traced UNet untouched.
+        from coremltools.optimize.coreml import (
+            OpPalettizerConfig,
+            OptimizationConfig,
+            palettize_weights,
+        )
+
+        nbits = int(quantize_nbits)
+        logger.info(f"Palettizing UNet weights to {nbits}-bit (kmeans)..")
+        t0 = time.time()
+        cfg = OptimizationConfig(
+            global_config=OpPalettizerConfig(mode="kmeans", nbits=nbits)
+        )
+        coreml_unet = palettize_weights(coreml_unet, config=cfg)
+        logger.info(f"Palettization took {time.time() - t0:.1f}s")
+
     coreml_unet.save(unet_out_path)
     logger.info(f"Saved unet into {unet_out_path}")
 
@@ -316,47 +280,43 @@ def convert(
     batch_size: int = 1,
     sample_size: tuple[int, int] = (64, 64),
     controlnet_support: bool = False,
-    lora_weights: list[tuple[Union[str, os.PathLike], float]] = None,
-    attn_impl: str = AttentionImplementations.SPLIT_EINSUM.name,
+    lora_weights: list[tuple[str | os.PathLike, float]] = None,
+    attn_impl: str = ATTENTION_IMPLEMENTATIONS[0],
     config_path: str = None,
+    quantize_nbits: str = "none",
 ):
     if os.path.exists(unet_out_path):
         logger.info(f"Found existing model at {unet_out_path}! Skipping..")
         return
 
-    python_coreml_stable_diffusion.unet.ATTENTION_IMPLEMENTATION_IN_EFFECT = (
-        AttentionImplementations(attn_impl)
-    )
-
-    ref_pipe = get_pipeline(ckpt_path, config_path, model_version)
+    if attn_impl not in ATTENTION_IMPLEMENTATIONS:
+        raise ValueError(
+            f"Unsupported attention implementation {attn_impl!r}. "
+            f"Expected one of {ATTENTION_IMPLEMENTATIONS}."
+        )
+    ref_unet = load_unet(ckpt_path, config_path)
 
     for i, lora_weight in enumerate(lora_weights or []):
         lora_path, strength = lora_weight
         adapter_name = f"lora_{i}"
-        ref_pipe.load_lora_weights(lora_path, adapter_name=adapter_name)
-        ref_pipe.set_adapters([adapter_name], adapter_weights=[strength])
-        ref_pipe.fuse_lora()
+        ref_unet.load_lora_adapter(lora_path, adapter_name=adapter_name)
+        ref_unet.set_adapters([adapter_name], weights=[strength])
+        ref_unet.fuse_lora()
 
     convert_unet(
-        ref_pipe,
+        ref_unet,
         model_version,
         unet_out_path,
         batch_size,
         sample_size,
         controlnet_support,
+        attention_implementation=attn_impl,
+        quantize_nbits=quantize_nbits,
     )
 
 
-def get_pipeline(ckpt_path, config_path, model_version):
-    pipe_cls = MODEL_TYPE_TO_PIPE_CLS[model_version]
-    ref_pipe = pipe_cls.from_single_file(ckpt_path, original_config_file=config_path)
-    return ref_pipe
-
-
-def compile_model(out_path, out_name, submodule_name):
-    # Compile the model
-    target_path = compile_coreml_model(
-        out_path, get_folder_paths(submodule_name)[0], f"{out_name}_{submodule_name}"
+def load_unet(ckpt_path, config_path):
+    return UNet2DConditionModel.from_single_file(
+        ckpt_path,
+        original_config=config_path,
     )
-    logger.info(f"Compiled {out_path} to {target_path}")
-    return target_path

coreml_suite/core/__init__.py

@@ -0,0 +1,10 @@
+"""Framework-free pure-logic core of ComfyUI-CoreMLSuite.
+
+Modules under this package must NOT import `comfy`, `coremltools`,
+`python_coreml_stable_diffusion`, `folder_paths`, `nodes`, or any other
+ComfyUI / Apple runtime. Only `numpy` and `torch` are allowed.
+
+The thin adapters in `coreml_suite.{latents,controlnet,models}` keep the
+old public import paths working so `coreml_suite/nodes.py` and downstream
+ComfyUI workflows are unchanged.
+"""

coreml_suite/core/controlnet.py

@@ -0,0 +1,67 @@
+"""Pure helpers around the ControlNet residual inputs of the Core ML UNet.
+
+Re-exported by coreml_suite.controlnet. Characterization tests cover
+shapes, dtype (fp16), and zero-fill fallback.
+"""
+from itertools import chain
+from math import ceil
+
+import numpy as np
+import torch
+
+from coreml_suite.core.latents import chunk_batch
+
+
+def expand_inputs(inputs):
+    expanded = inputs.copy()
+    for k, v in inputs.items():
+        if isinstance(v, np.ndarray):
+            expanded[k] = np.concatenate([v] * 2) if v.shape[0] == 1 else v
+        elif isinstance(v, torch.Tensor):
+            expanded[k] = torch.cat([v] * 2) if v.shape[0] == 1 else v
+        elif isinstance(v, list):
+            expanded[k] = v * 2 if len(v) == 1 else v
+        elif isinstance(v, dict):
+            expand_inputs(v)
+    return expanded
+
+
+def extract_residual_kwargs(expected_inputs, control):
+    if "additional_residual_0" not in expected_inputs.keys():
+        return {}
+    if control is None:
+        return no_control(expected_inputs)
+
+    residual_kwargs = {
+        "additional_residual_{}".format(i): r.cpu().numpy().astype(np.float16)
+        for i, r in enumerate(chain(control["output"], control["middle"]))
+    }
+    return residual_kwargs
+
+
+def no_control(expected_inputs):
+    shapes_dict = {
+        k: v["shape"] for k, v in expected_inputs.items() if k.startswith("additional")
+    }
+    residual_kwargs = {
+        k: torch.zeros(*shape).cpu().numpy().astype(dtype=np.float16)
+        for k, shape in shapes_dict.items()
+    }
+    return residual_kwargs
+
+
+def chunk_control(cn, target_size):
+    if cn is None:
+        return [None] * target_size
+
+    num_chunks = ceil(cn["output"][0].shape[0] / target_size)
+
+    out = [{"output": [], "middle": []} for _ in range(num_chunks)]
+
+    for k, v in cn.items():
+        for i, x in enumerate(v):
+            chunks = chunk_batch(x, (target_size, *x.shape[1:]))
+            for j, chunk in enumerate(chunks):
+                out[j][k].append(chunk)
+
+    return out

coreml_suite/core/inputs.py

@@ -0,0 +1,111 @@
+"""Pure transform from torch sampler inputs to Core ML UNet kwargs.
+
+Characterization tests cover SD1.5 / SDXL base / SDXL refiner / LCM
+variants and the chunked-batch fan-out.
+"""
+import numpy as np
+import torch
+
+from coreml_suite.core.controlnet import extract_residual_kwargs, chunk_control
+from coreml_suite.core.latents import chunk_batch
+
+
+class CoreMLInputs:
+    def __init__(self, x, t, context, control, **kwargs):
+        self.x = x
+        self.t = t
+        self.context = context
+        self.control = control
+        self.time_ids = kwargs.get("time_ids")
+        self.text_embeds = kwargs.get("text_embeds")
+        self.ts_cond = kwargs.get("timestep_cond")
+
+    def coreml_kwargs(self, expected_inputs):
+        sample = self.x.cpu().numpy().astype(np.float16)
+
+        context = self.context.cpu().numpy().astype(np.float16)
+
+        t = self.t.cpu().numpy().astype(np.float16)
+
+        model_input_kwargs = {
+            "sample": sample,
+            "encoder_hidden_states": context,
+            "timestep": t,
+        }
+        residual_kwargs = extract_residual_kwargs(expected_inputs, self.control)
+        model_input_kwargs |= residual_kwargs
+
+        # LCM
+        if self.ts_cond is not None:
+            model_input_kwargs["timestep_cond"] = (
+                self.ts_cond.cpu().numpy().astype(np.float16)
+            )
+
+        # SDXL
+        if "text_embeds" in expected_inputs:
+            model_input_kwargs["text_embeds"] = (
+                self.text_embeds.cpu().numpy().astype(np.float16)
+            )
+        if "time_ids" in expected_inputs:
+            model_input_kwargs["time_ids"] = (
+                self.time_ids.cpu().numpy().astype(np.float16)
+            )
+
+        return model_input_kwargs
+
+    def chunks(self, expected_inputs):
+        sample_shape = expected_inputs["sample"]["shape"]
+        timestep_shape = expected_inputs["timestep"]["shape"]
+        context_shape = expected_inputs["encoder_hidden_states"]["shape"]
+
+        chunked_x = chunk_batch(self.x, sample_shape)
+        ts = list(torch.full((len(chunked_x), timestep_shape[0]), self.t[0]))
+        chunked_context = chunk_batch(self.context, context_shape)
+
+        chunked_control = [None] * len(chunked_x)
+        if self.control is not None:
+            chunked_control = chunk_control(self.control, sample_shape[0])
+
+        chunked_ts_cond = [None] * len(chunked_x)
+        if self.ts_cond is not None:
+            ts_cond_shape = expected_inputs["timestep_cond"]["shape"]
+            chunked_ts_cond = chunk_batch(self.ts_cond, ts_cond_shape)
+
+        chunked_time_ids = [None] * len(chunked_x)
+        if expected_inputs.get("time_ids") is not None:
+            time_ids_shape = expected_inputs["time_ids"]["shape"]
+            if self.time_ids is None:
+                self.time_ids = torch.zeros(len(chunked_x), *time_ids_shape[1:]).to(
+                    self.x.device
+                )
+            chunked_time_ids = chunk_batch(self.time_ids, time_ids_shape)
+
+        chunked_text_embeds = [None] * len(chunked_x)
+        if expected_inputs.get("text_embeds") is not None:
+            text_embeds_shape = expected_inputs["text_embeds"]["shape"]
+            if self.text_embeds is None:
+                self.text_embeds = torch.zeros(
+                    len(chunked_x), *text_embeds_shape[1:]
+                ).to(self.x.device)
+            chunked_text_embeds = chunk_batch(self.text_embeds, text_embeds_shape)
+
+        return [
+            CoreMLInputs(
+                x,
+                t,
+                context,
+                control,
+                timestep_cond=ts_cond,
+                time_ids=time_ids,
+                text_embeds=text_embeds,
+            )
+            for x, t, context, control, ts_cond, time_ids, text_embeds in zip(
+                chunked_x,
+                ts,
+                chunked_context,
+                chunked_control,
+                chunked_ts_cond,
+                chunked_time_ids,
+                chunked_text_embeds,
+            )
+        ]

coreml_suite/core/latents.py

@@ -0,0 +1,42 @@
+"""Pure batch-chunking helpers for Core ML's fixed-shape UNet inputs.
+
+Re-exported by coreml_suite.latents. Characterization tests cover the
+contract (padding-zero regions, truncation in merge_chunks,
+identity-passthrough when shape already matches).
+"""
+import torch
+
+
+def chunk_batch(input_tensor, target_shape):
+    if input_tensor.shape == target_shape:
+        return [input_tensor]
+
+    batch_size = input_tensor.shape[0]
+    target_batch_size = target_shape[0]
+
+    num_chunks = batch_size // target_batch_size
+    if num_chunks == 0:
+        padding = torch.zeros(target_batch_size - batch_size, *target_shape[1:]).to(
+            input_tensor.device
+        )
+        return [torch.cat((input_tensor, padding), dim=0)]
+
+    mod = batch_size % target_batch_size
+    if mod != 0:
+        chunks = list(torch.chunk(input_tensor[:-mod], num_chunks))
+        padding = torch.zeros(target_batch_size - mod, *target_shape[1:]).to(
+            input_tensor.device
+        )
+        padded = torch.cat((input_tensor[-mod:], padding), dim=0)
+        chunks.append(padded)
+        return chunks
+
+    chunks = list(torch.chunk(input_tensor, num_chunks))
+    return chunks
+
+
+def merge_chunks(chunks, orig_shape):
+    merged = torch.cat(chunks, dim=0)
+    if merged.shape == orig_shape:
+        return merged
+    return merged[: orig_shape[0]]

coreml_suite/core/naming.py

@@ -0,0 +1,68 @@
+"""Pure out_name composition for the Core ML UNet artifact.
+
+Extracted from CoreMLConverter.convert so the filename contract
+can be tested + reused without instantiating the node. The string is the
+cache key: every workflow that references a converted .mlpackage depends
+on it staying byte-for-byte identical.
+"""
+from typing import Iterable, Tuple
+
+ATTN_SUFFIX = {
+    "SPLIT_EINSUM": "se",
+    "SPLIT_EINSUM_V2": "se2",
+    "ORIGINAL": "orig",
+}
+
+# Palettization bits. "none" = no quantization (default; keeps the
+# unquantized filename intact so existing workflows still resolve their
+# cached .mlpackage). Numeric values append a `_q<bits>` suffix.
+QUANT_NBITS_VALUES = ("none", "8", "6", "4")
+
+
+def compose_out_name(
+    *,
+    ckpt_name: str,
+    batch_size: int,
+    width: int,
+    height: int,
+    controlnet_support: bool,
+    attention_implementation: str,
+    lora_names: Iterable[str] = (),
+    quantize_nbits: str = "none",
+) -> str:
+    """Build the .mlpackage stem from convert() parameters.
+
+    Locked behaviour (characterization tests):
+      - first '.' in ckpt_name wins (`a.b.c.safetensors` -> `a`)
+      - spaces collapse to underscores
+      - LoRA names are taken stem-only, sorted, joined with '_' and
+        prefixed with '_' when present (caller is expected to pass a
+        sorted list; we sort defensively)
+      - controlnet adds `_cn`
+      - attn suffix is `_se` | `_se2` | `_orig`
+
+    Quantization:
+      - quantize_nbits "none" (default) appends nothing — existing
+        unquantized .mlpackages keep the old filename
+      - "4" / "6" / "8" appends `_q<bits>` after the attn suffix
+    """
+    if quantize_nbits not in QUANT_NBITS_VALUES:
+        raise ValueError(
+            f"quantize_nbits={quantize_nbits!r} not in {QUANT_NBITS_VALUES}"
+        )
+    stem = ckpt_name.split(".")[0]
+    sorted_names = sorted(lora_names)
+    lora_str = "_" + "_".join(name.split(".")[0] for name in sorted_names) if sorted_names else ""
+    cn_suffix = "_cn" if controlnet_support else ""
+    attn_suffix = "_" + ATTN_SUFFIX[attention_implementation]
+    quant_suffix = f"_q{quantize_nbits}" if quantize_nbits != "none" else ""
+    out_name = (
+        f"{stem}{lora_str}_{batch_size}x{width}x{height}"
+        f"{cn_suffix}{attn_suffix}{quant_suffix}"
+    )
+    return out_name.replace(" ", "_")
+
+
+def lora_names_from_params(lora_params: Iterable[Tuple[str, float]]) -> list[str]:
+    """Mirror the sort applied inside CoreMLConverter.convert."""
+    return [name for name, _ in sorted(lora_params, key=lambda pair: pair[0])]

coreml_suite/core/sdxl.py

@@ -0,0 +1,91 @@
+"""Pure SDXL detection + time_ids/text_embeds assembly.
+
+The framework-coupled adapter `add_sdxl_model_options` lives in models.py
+and delegates the math here. Characterization tests cover base (len 6) vs
+refiner (len 5) and the closure free-vars produced by
+`sdxl_model_function_wrapper`.
+"""
+import torch
+
+
+def is_sdxl(coreml_model):
+    return (
+        "time_ids" in coreml_model.expected_inputs
+        and "text_embeds" in coreml_model.expected_inputs
+    )
+
+
+def is_sdxl_base(coreml_model):
+    return (
+        is_sdxl(coreml_model)
+        and coreml_model.expected_inputs["time_ids"]["shape"][1] == 6
+    )
+
+
+def is_sdxl_refiner(coreml_model):
+    return (
+        is_sdxl(coreml_model)
+        and coreml_model.expected_inputs["time_ids"]["shape"][1] == 5
+    )
+
+
+def build_sdxl_time_ids(pos_dict, neg_dict, *, is_base: bool, is_refiner: bool):
+    """Compose the (2, N) time_ids tensor for the SDXL Core ML UNet.
+
+    - base: N=6  -> [h, w, crop_h, crop_w, target_h, target_w]
+    - refiner: N=5 -> [h, w, crop_h, crop_w, aesthetic_score]
+    - neither: N=4 -> [h, w, crop_h, crop_w] (edge case kept for parity)
+    """
+    pos_time_ids = [
+        pos_dict.get("height", 768),
+        pos_dict.get("width", 768),
+        pos_dict.get("crop_h", 0),
+        pos_dict.get("crop_w", 0),
+    ]
+    neg_time_ids = [
+        neg_dict.get("height", 768),
+        neg_dict.get("width", 768),
+        neg_dict.get("crop_h", 0),
+        neg_dict.get("crop_w", 0),
+    ]
+
+    if is_base:
+        pos_time_ids += [
+            pos_dict.get("target_height", 768),
+            pos_dict.get("target_width", 768),
+        ]
+        neg_time_ids += [
+            neg_dict.get("target_height", 768),
+            neg_dict.get("target_width", 768),
+        ]
+
+    if is_refiner:
+        pos_time_ids += [pos_dict.get("aesthetic_score", 6)]
+        neg_time_ids += [neg_dict.get("aesthetic_score", 2.5)]
+
+    return torch.tensor([pos_time_ids, neg_time_ids])
+
+
+def build_sdxl_text_embeds(pos_pooled, neg_pooled):
+    """Concat pos then neg along the batch dim. Locked contract."""
+    return torch.cat((pos_pooled, neg_pooled))
+
+
+def sdxl_model_function_wrapper(time_ids, text_embeds, refiner=False):
+    def wrapper(model_function, params):
+        x = params["input"]
+        t = params["timestep"]
+        c = params["c"]
+
+        context = c.get("c_crossattn")
+
+        if context is None:
+            return torch.zeros_like(x)
+
+        if refiner and context is not None:
+            # converted refiner accepts only g clip
+            c["c_crossattn"] = context[:, :, 768:]
+
+        return model_function(x, t, **c, time_ids=time_ids, text_embeds=text_embeds)
+
+    return wrapper

coreml_suite/coreml_model.py

@@ -0,0 +1,42 @@
+import time
+
+import coremltools as ct
+
+from coreml_suite.logger import logger
+
+
+class CoreMLModel:
+    """Small runtime wrapper around coremltools.models.MLModel.
+
+    This keeps the inference path independent from apple/ml-stable-diffusion's
+    CoreMLModel wrapper while preserving the contract used by the sampler code:
+    ``expected_inputs`` and callable prediction.
+    """
+
+    def __init__(self, model_path, compute_unit):
+        self.model_path = model_path
+        self.compute_unit = self._compute_unit(compute_unit)
+
+        logger.info(f"Loading {model_path} to {self.compute_unit.name}")
+        start = time.time()
+        self.model = ct.models.MLModel(model_path, compute_units=self.compute_unit)
+        logger.info(f"Loading {model_path} took {time.time() - start:.1f} seconds")
+
+        self.expected_inputs = self._expected_inputs()
+
+    def __call__(self, **kwargs):
+        return self.model.predict(kwargs)
+
+    @staticmethod
+    def _compute_unit(compute_unit):
+        if isinstance(compute_unit, ct.ComputeUnit):
+            return compute_unit
+        return ct.ComputeUnit[compute_unit]
+
+    def _expected_inputs(self):
+        return {
+            feature.name: {
+                "shape": tuple(feature.type.multiArrayType.shape),
+            }
+            for feature in self.model.get_spec().description.input
+        }

coreml_suite/latents.py

@@ -1,36 +1,4 @@
-import torch
+"""Compatibility shim — re-exports from coreml_suite.core.latents."""
+from coreml_suite.core.latents import chunk_batch, merge_chunks
 
-
-def chunk_batch(input_tensor, target_shape):
-    if input_tensor.shape == target_shape:
-        return [input_tensor]
-
-    batch_size = input_tensor.shape[0]
-    target_batch_size = target_shape[0]
-
-    num_chunks = batch_size // target_batch_size
-    if num_chunks == 0:
-        padding = torch.zeros(target_batch_size - batch_size, *target_shape[1:]).to(
-            input_tensor.device
-        )
-        return [torch.cat((input_tensor, padding), dim=0)]
-
-    mod = batch_size % target_batch_size
-    if mod != 0:
-        chunks = list(torch.chunk(input_tensor[:-mod], num_chunks))
-        padding = torch.zeros(target_batch_size - mod, *target_shape[1:]).to(
-            input_tensor.device
-        )
-        padded = torch.cat((input_tensor[-mod:], padding), dim=0)
-        chunks.append(padded)
-        return chunks
-
-    chunks = list(torch.chunk(input_tensor, num_chunks))
-    return chunks
-
-
-def merge_chunks(chunks, orig_shape):
-    merged = torch.cat(chunks, dim=0)
-    if merged.shape == orig_shape:
-        return merged
-    return merged[: orig_shape[0]]
+__all__ = ["chunk_batch", "merge_chunks"]

coreml_suite/lcm/converter.py

@@ -1,5 +1,4 @@
 import os
-import shutil
 import logging
 import time
 import gc
@@ -9,24 +8,20 @@ import torch
 from diffusers import UNet2DConditionModel, LCMScheduler
 from diffusers.loaders import LoraLoaderMixin
 
-from comfy.model_management import get_torch_device
-from coreml_suite.lcm.unet import UNet2DConditionModelLCM
+from coreml_suite.conversion.attention import apply_attention_implementation
+from coreml_suite.conversion.shapes import conv2d_output_shape
+from coreml_suite.conversion.unet import CoreMLUNetWrapper
+from coreml_suite.model_version import ModelVersion
 
-from transformers import CLIPTextModel
 import coremltools as ct
 
-from folder_paths import get_folder_paths
-
 logging.basicConfig()
 logger = logging.getLogger(__name__)
 logger.setLevel(logging.DEBUG)
 
 MODEL_VERSION = "SimianLuo/LCM_Dreamshaper_v7"
 MODEL_NAME = MODEL_VERSION.split("/")[-1] + "_4k"
-
-import python_coreml_stable_diffusion.unet as unet
-
-unet.ATTENTION_IMPLEMENTATION_IN_EFFECT = unet.AttentionImplementations.SPLIT_EINSUM
+TEXT_TOKEN_SEQUENCE_LENGTH = 77
 
 
 def get_unets():
@@ -37,31 +32,27 @@ def get_unets():
         low_cpu_mem_usage=False,
     )
 
-    cml_unet = UNet2DConditionModelLCM.from_config(ref_unet.config).eval()
-    cml_unet.load_state_dict(ref_unet.state_dict(), strict=False)
+    cml_unet = CoreMLUNetWrapper(
+        apply_attention_implementation(ref_unet.eval(), "SPLIT_EINSUM"),
+        ModelVersion.LCM,
+    )
 
     return cml_unet, ref_unet
 
 
 def get_encoder_hidden_states_shape(unet_config, batch_size):
-    text_encoder = CLIPTextModel.from_pretrained(
-        MODEL_VERSION, subfolder="text_encoder"
-    )
-
-    text_token_sequence_length = text_encoder.config.max_position_embeddings
-    hidden_size = (text_encoder.config.hidden_size,)
-
     encoder_hidden_states_shape = (
         batch_size,
-        unet_config.cross_attention_dim or hidden_size,
-        1,
-        text_token_sequence_length,
+        TEXT_TOKEN_SEQUENCE_LENGTH,
+        unet_config.cross_attention_dim,
     )
 
     return encoder_hidden_states_shape
 
 
 def get_scheduler():
+    from comfy.model_management import get_torch_device
+
     scheduler = LCMScheduler.from_pretrained(MODEL_VERSION, subfolder="scheduler")
     scheduler.set_timesteps(50, get_torch_device(), 50)
     return scheduler
@@ -117,29 +108,14 @@ def convert_to_coreml(
 
 
 def get_out_path(submodule_name, model_name):
+    from folder_paths import get_folder_paths
+
     fname = f"{model_name}_{submodule_name}.mlpackage"
     unet_path = get_folder_paths(submodule_name)[0]
     out_path = os.path.join(unet_path, fname)
     return out_path
 
 
-def compile_coreml_model(source_model_path, output_dir, final_name):
-    """Compiles Core ML models using the coremlcompiler utility from Xcode toolchain"""
-    target_path = os.path.join(output_dir, f"{final_name}.mlmodelc")
-    if os.path.exists(target_path):
-        logger.warning(f"Found existing compiled model at {target_path}! Skipping..")
-        return target_path
-
-    logger.info(f"Compiling {source_model_path}")
-    source_model_name = os.path.basename(os.path.splitext(source_model_path)[0])
-
-    os.system(f"xcrun coremlcompiler compile {source_model_path} {output_dir}")
-    compiled_output = os.path.join(output_dir, f"{source_model_name}.mlmodelc")
-    shutil.move(compiled_output, target_path)
-
-    return target_path
-
-
 def get_sample_input(batch_size, encoder_hidden_states_shape, sample_shape, scheduler):
     sample_unet_inputs = dict(
         [
@@ -165,15 +141,13 @@ def get_unet_inputs_spec(sample_unet_inputs):
 
 
 def add_cnet_support(sample_shape, reference_unet):
-    from python_coreml_stable_diffusion.unet import calculate_conv2d_output_shape
-
     additional_residuals_shapes = []
 
     batch_size = sample_shape[0]
     h, w = sample_shape[2:]
 
     # conv_in
-    out_h, out_w = calculate_conv2d_output_shape(
+    out_h, out_w = conv2d_output_shape(
         h,
         w,
         reference_unet.conv_in,
@@ -190,9 +164,7 @@ def add_cnet_support(sample_shape, reference_unet):
         ]
         if hasattr(down_block, "downsamplers") and down_block.downsamplers is not None:
             for downsampler in down_block.downsamplers:
-                out_h, out_w = calculate_conv2d_output_shape(
-                    out_h, out_w, downsampler.conv
-                )
+                out_h, out_w = conv2d_output_shape(out_h, out_w, downsampler.conv)
             additional_residuals_shapes.append(
                 (
                     batch_size,
@@ -272,15 +244,6 @@ def convert(
     logger.info(f"Saved unet into {out_path}")
 
 
-def compile_model(out_path, out_name):
-    # Compile the model
-    target_path = compile_coreml_model(
-        out_path, get_folder_paths("unet")[0], f"{out_name}_unet"
-    )
-    logger.info(f"Compiled {out_path} to {target_path}")
-    return target_path
-
-
 if __name__ == "__main__":
     h = 512
     w = 512
@@ -294,4 +257,3 @@ if __name__ == "__main__":
     out_path = get_out_path("unet", f"{out_name}")
     if not os.path.exists(out_path):
         convert(out_path=out_path, sample_size=sample_size, batch_size=batch_size)
-    compile_model(out_path=out_path, out_name=out_name)

coreml_suite/lcm/nodes.py

@@ -1,10 +1,9 @@
 import os
 
 from coremltools import ComputeUnit
-from python_coreml_stable_diffusion.coreml_model import CoreMLModel
 
 from coreml_suite import COREML_NODE
-from coreml_suite.lcm import converter as lcm_converter
+from coreml_suite.coreml_model import CoreMLModel
 
 
 class COREML_CONVERT_LCM(COREML_NODE):
@@ -48,6 +47,8 @@ class COREML_CONVERT_LCM(COREML_NODE):
         The converted model is also saved to "models/unet" directory and
         can be loaded with the "LCMCoreMLLoaderUNet" node.
         """
+        from coreml_suite.lcm import converter as lcm_converter
+
         h = height
         w = width
         sample_size = (h // 8, w // 8)
@@ -65,6 +66,5 @@ class COREML_CONVERT_LCM(COREML_NODE):
                 batch_size=batch_size,
                 controlnet_support=controlnet_support,
             )
-        target_path = lcm_converter.compile_model(out_path=out_path, out_name=out_name)
 
-        return (CoreMLModel(target_path, compute_unit, "compiled"),)
+        return (CoreMLModel(out_path, compute_unit),)

coreml_suite/lcm/unet.py

@@ -1,5 +1,5 @@
-from overrides import overrides
-from python_coreml_stable_diffusion.unet import UNet2DConditionModel, TimestepEmbedding
+from diffusers import UNet2DConditionModel
+from diffusers.models.embeddings import TimestepEmbedding
 
 
 class UNet2DConditionModelLCM(UNet2DConditionModel):
@@ -17,7 +17,6 @@ class UNet2DConditionModelLCM(UNet2DConditionModel):
         )
         self.time_embedding = time_embedding
 
-    @overrides(check_signature=False)
     def forward(
         self,
         sample,

coreml_suite/model_version.py

@@ -0,0 +1,8 @@
+from enum import Enum
+
+
+class ModelVersion(Enum):
+    SD15 = "sd15"
+    SDXL = "sdxl"
+    SDXL_REFINER = "sdxl_refiner"
+    LCM = "lcm"

coreml_suite/models.py

@@ -1,15 +1,44 @@
-import numpy as np
+"""Framework-coupled glue between Core ML UNets and ComfyUI's sampler stack.
+
+Pure math (CoreMLInputs, SDXL detection, time_ids/text_embeds assembly,
+sdxl_model_function_wrapper) lives in coreml_suite.core.*.
+This module is what touches comfy.*: model_base, ModelPatcher, the
+diffusion_model wrapper, and the maintainer-facing add_sdxl_model_options
+adapter.
+"""
 import torch
 
 from comfy import model_base
 from comfy.model_management import get_torch_device
 from comfy.model_patcher import ModelPatcher
+
 from coreml_suite.config import get_model_config, ModelVersion
-from coreml_suite.controlnet import extract_residual_kwargs, chunk_control
-from coreml_suite.latents import chunk_batch, merge_chunks
+from coreml_suite.core.inputs import CoreMLInputs
+from coreml_suite.core.latents import merge_chunks
+from coreml_suite.core.sdxl import (
+    build_sdxl_text_embeds,
+    build_sdxl_time_ids,
+    is_sdxl,
+    is_sdxl_base,
+    is_sdxl_refiner,
+    sdxl_model_function_wrapper,
+)
 from coreml_suite.lcm.utils import is_lcm
 from coreml_suite.logger import logger
 
+__all__ = [
+    "CoreMLInputs",
+    "CoreMLModelWrapper",
+    "CoreMLModelWrapperLCM",
+    "add_sdxl_model_options",
+    "get_latent_image",
+    "get_model_patcher",
+    "is_sdxl",
+    "is_sdxl_base",
+    "is_sdxl_refiner",
+    "sdxl_model_function_wrapper",
+]
+
 
 class CoreMLModelWrapper:
     def __init__(self, coreml_model):
@@ -68,204 +97,27 @@ class CoreMLModelWrapperLCM(CoreMLModelWrapper):
         self.config = None
 
 
-class CoreMLInputs:
-    def __init__(self, x, t, context, control, **kwargs):
-        self.x = x
-        self.t = t
-        self.context = context
-        self.control = control
-        self.time_ids = kwargs.get("time_ids")
-        self.text_embeds = kwargs.get("text_embeds")
-        self.ts_cond = kwargs.get("timestep_cond")
-
-    def coreml_kwargs(self, expected_inputs):
-        sample = self.x.cpu().numpy().astype(np.float16)
-
-        context = self.context.cpu().numpy().astype(np.float16)
-        context = context.transpose(0, 2, 1)[:, :, None, :]
-
-        t = self.t.cpu().numpy().astype(np.float16)
-
-        model_input_kwargs = {
-            "sample": sample,
-            "encoder_hidden_states": context,
-            "timestep": t,
-        }
-        residual_kwargs = extract_residual_kwargs(expected_inputs, self.control)
-        model_input_kwargs |= residual_kwargs
-
-        # LCM
-        if self.ts_cond is not None:
-            model_input_kwargs["timestep_cond"] = (
-                self.ts_cond.cpu().numpy().astype(np.float16)
-            )
-
-        # SDXL
-        if "text_embeds" in expected_inputs:
-            model_input_kwargs["text_embeds"] = (
-                self.text_embeds.cpu().numpy().astype(np.float16)
-            )
-        if "time_ids" in expected_inputs:
-            model_input_kwargs["time_ids"] = (
-                self.time_ids.cpu().numpy().astype(np.float16)
-            )
-
-        return model_input_kwargs
-
-    def chunks(self, expected_inputs):
-        sample_shape = expected_inputs["sample"]["shape"]
-        timestep_shape = expected_inputs["timestep"]["shape"]
-        hidden_shape = expected_inputs["encoder_hidden_states"]["shape"]
-        context_shape = (hidden_shape[0], hidden_shape[3], hidden_shape[1])
-
-        chunked_x = chunk_batch(self.x, sample_shape)
-        ts = list(torch.full((len(chunked_x), timestep_shape[0]), self.t[0]))
-        chunked_context = chunk_batch(self.context, context_shape)
-
-        chunked_control = [None] * len(chunked_x)
-        if self.control is not None:
-            chunked_control = chunk_control(self.control, sample_shape[0])
-
-        chunked_ts_cond = [None] * len(chunked_x)
-        if self.ts_cond is not None:
-            ts_cond_shape = expected_inputs["timestep_cond"]["shape"]
-            chunked_ts_cond = chunk_batch(self.ts_cond, ts_cond_shape)
-
-        chunked_time_ids = [None] * len(chunked_x)
-        if expected_inputs.get("time_ids") is not None:
-            time_ids_shape = expected_inputs["time_ids"]["shape"]
-            if self.time_ids is None:
-                self.time_ids = torch.zeros(len(chunked_x), *time_ids_shape[1:]).to(
-                    self.x.device
-                )
-            chunked_time_ids = chunk_batch(self.time_ids, time_ids_shape)
-
-        chunked_text_embeds = [None] * len(chunked_x)
-        if expected_inputs.get("text_embeds") is not None:
-            text_embeds_shape = expected_inputs["text_embeds"]["shape"]
-            if self.text_embeds is None:
-                self.text_embeds = torch.zeros(
-                    len(chunked_x), *text_embeds_shape[1:]
-                ).to(self.x.device)
-            chunked_text_embeds = chunk_batch(self.text_embeds, text_embeds_shape)
-
-        return [
-            CoreMLInputs(
-                x,
-                t,
-                context,
-                control,
-                timestep_cond=ts_cond,
-                time_ids=time_ids,
-                text_embeds=text_embeds,
-            )
-            for x, t, context, control, ts_cond, time_ids, text_embeds in zip(
-                chunked_x,
-                ts,
-                chunked_context,
-                chunked_control,
-                chunked_ts_cond,
-                chunked_time_ids,
-                chunked_text_embeds,
-            )
-        ]
-
-
-def is_sdxl(coreml_model):
-    return (
-        "time_ids" in coreml_model.expected_inputs
-        and "text_embeds" in coreml_model.expected_inputs
-    )
-
-
-def is_sdxl_base(coreml_model):
-    return (
-        is_sdxl(coreml_model)
-        and coreml_model.expected_inputs["time_ids"]["shape"][1] == 6
-    )
-
-
-def is_sdxl_refiner(coreml_model):
-    return (
-        is_sdxl(coreml_model)
-        and coreml_model.expected_inputs["time_ids"]["shape"][1] == 5
-    )
-
-
-def sdxl_model_function_wrapper(time_ids, text_embeds, refiner=False):
-    def wrapper(model_function, params):
-        x = params["input"]
-        t = params["timestep"]
-        c = params["c"]
-
-        context = c.get("c_crossattn")
-
-        if context is None:
-            return torch.zeros_like(x)
-
-        if refiner and context is not None:
-            # converted refiner accepts only g clip
-            c["c_crossattn"] = context[:, :, 768:]
-
-        return model_function(x, t, **c, time_ids=time_ids, text_embeds=text_embeds)
-
-    return wrapper
-
-
 def add_sdxl_model_options(model_patcher, positive, negative):
     mp = model_patcher.clone()
 
     pos_dict = positive[0][1]
     neg_dict = negative[0][1]
 
-    pos_pooled = pos_dict["pooled_output"]
-    neg_pooled = neg_dict["pooled_output"]
-
-    pos_time_ids = [
-        pos_dict.get("height", 768),
-        pos_dict.get("width", 768),
-        pos_dict.get("crop_h", 0),
-        pos_dict.get("crop_w", 0),
-    ]
-
-    neg_time_ids = [
-        neg_dict.get("height", 768),
-        neg_dict.get("width", 768),
-        neg_dict.get("crop_h", 0),
-        neg_dict.get("crop_w", 0),
-    ]
-
-    if model_patcher.model.diffusion_model.is_sdxl_base:
-        pos_time_ids += [
-            pos_dict.get("target_height", 768),
-            pos_dict.get("target_width", 768),
-        ]
-
-        neg_time_ids += [
-            neg_dict.get("target_height", 768),
-            neg_dict.get("target_width", 768),
-        ]
-
+    is_base = model_patcher.model.diffusion_model.is_sdxl_base
     is_refiner = model_patcher.model.diffusion_model.is_sdxl_refiner
-    if is_refiner:
-        pos_time_ids += [
-            pos_dict.get("aesthetic_score", 6),
-        ]
 
-        neg_time_ids += [
-            neg_dict.get("aesthetic_score", 2.5),
-        ]
+    time_ids = build_sdxl_time_ids(
+        pos_dict, neg_dict, is_base=is_base, is_refiner=is_refiner
+    )
+    text_embeds = build_sdxl_text_embeds(
+        pos_dict["pooled_output"], neg_dict["pooled_output"]
+    )
 
-    time_ids = torch.tensor([pos_time_ids, neg_time_ids])
-    text_embeds = torch.cat((pos_pooled, neg_pooled))
-
-    model_options = {
+    mp.model_options |= {
         "model_function_wrapper": sdxl_model_function_wrapper(
             time_ids, text_embeds, is_refiner
         ),
     }
-    mp.model_options |= model_options
-
     return mp
 
 

coreml_suite/nodes.py

@@ -1,15 +1,19 @@
 import os
 
 from coremltools import ComputeUnit
-from python_coreml_stable_diffusion.coreml_model import CoreMLModel
-from python_coreml_stable_diffusion.unet import AttentionImplementations
 
 import folder_paths
 from coreml_suite import COREML_NODE
-from coreml_suite import converter
-from coreml_suite.config import ModelVersion
+from coreml_suite.attention import ATTENTION_IMPLEMENTATIONS
+from coreml_suite.coreml_model import CoreMLModel
+from coreml_suite.core.naming import (
+    QUANT_NBITS_VALUES,
+    compose_out_name,
+    lora_names_from_params,
+)
 from coreml_suite.lcm.utils import add_lcm_model_options, lcm_patch, is_lcm
 from coreml_suite.logger import logger
+from coreml_suite.model_version import ModelVersion
 from nodes import KSampler, LoraLoader, KSamplerAdvanced
 
 from coreml_suite.models import (
@@ -163,7 +167,7 @@ class CoreMLLoader(COREML_NODE):
 
     @classmethod
     def coreml_filenames(cls):
-        extensions = (".mlmodelc", ".mlpackage")
+        extensions = (".mlpackage",)
         all_paths = folder_paths.get_filename_list_(cls.PACKAGE_DIRNAME)[1]
         coreml_paths = folder_paths.filter_files_extensions(all_paths, extensions)
 
@@ -174,9 +178,7 @@ class CoreMLLoader(COREML_NODE):
 
         coreml_path = self.coreml_filenames()[coreml_name]
 
-        sources = "compiled" if coreml_name.endswith(".mlmodelc") else "packages"
-
-        return (CoreMLModel(coreml_path, compute_unit, sources),)
+        return (CoreMLModel(coreml_path, compute_unit),)
 
 
 class CoreMLLoaderUNet(CoreMLLoader):
@@ -223,15 +225,11 @@ class CoreMLConverter(COREML_NODE):
                         ModelVersion.SDXL.name,
                     ],
                 ),
-                "height": ("INT", {"default": 512, "min": 256, "max": 2048, "step": 8}),
-                "width": ("INT", {"default": 512, "min": 256, "max": 2048, "step": 8}),
+                "height": ("INT", {"default": 512, "min": 8, "step": 8}),
+                "width": ("INT", {"default": 512, "min": 8, "step": 8}),
                 "batch_size": ("INT", {"default": 1, "min": 1, "max": 64}),
                 "attention_implementation": (
-                    [
-                        AttentionImplementations.SPLIT_EINSUM.name,
-                        AttentionImplementations.SPLIT_EINSUM_V2.name,
-                        AttentionImplementations.ORIGINAL.name,
-                    ],
+                    list(ATTENTION_IMPLEMENTATIONS),
                 ),
                 "compute_unit": (
                     [
@@ -244,6 +242,12 @@ class CoreMLConverter(COREML_NODE):
                 "controlnet_support": ("BOOLEAN", {"default": False}),
             },
             "optional": {
+                # k-means weight palettization. Kept optional so workflows
+                # that omit it still validate — ComfyUI rejects a prompt that
+                # omits any `required` input. When omitted it defaults to
+                # "none", identical to unquantized behavior and filename, so
+                # existing cached .mlpackages still resolve.
+                "quantize_nbits": (list(QUANT_NBITS_VALUES), {"default": "none"}),
                 "lora_params": ("LORA_PARAMS",),
             },
         }
@@ -262,6 +266,7 @@ class CoreMLConverter(COREML_NODE):
         attention_implementation,
         compute_unit,
         controlnet_support,
+        quantize_nbits="none",
         lora_params=None,
     ):
         """Converts a LCM model to Core ML.
@@ -288,24 +293,17 @@ class CoreMLConverter(COREML_NODE):
         h = height
         w = width
         sample_size = (h // 8, w // 8)
-        batch_size = batch_size
-        cn_support_str = "_cn" if controlnet_support else ""
-        lora_str = (
-            "_" + "_".join(lora_param[0].split(".")[0] for lora_param in lora_params)
-            if lora_params
-            else ""
+        out_name = compose_out_name(
+            ckpt_name=ckpt_name,
+            batch_size=batch_size,
+            width=w,
+            height=h,
+            controlnet_support=controlnet_support,
+            attention_implementation=attention_implementation,
+            lora_names=lora_names_from_params(lora_params),
+            quantize_nbits=quantize_nbits,
         )
 
-        attn_str = (
-            "_"
-            + {"SPLIT_EINSUM": "se", "SPLIT_EINSUM_V2": "se2", "ORIGINAL": "orig"}[
-                attention_implementation
-            ]
-        )
-
-        out_name = f"{ckpt_name.split('.')[0]}{lora_str}_{batch_size}x{w}x{h}{cn_support_str}{attn_str}"
-        out_name = out_name.replace(" ", "_")
-
         logger.info(f"Converting {ckpt_name} to {out_name}")
         logger.info(f"Batch size: {batch_size}")
         logger.info(f"Width: {w}, Height: {h}")
@@ -317,6 +315,8 @@ class CoreMLConverter(COREML_NODE):
             for lora_param in lora_params:
                 logger.info(f"  {lora_param[0]} - strength: {lora_param[1]}")
 
+        from coreml_suite import converter
+
         unet_out_path = converter.get_out_path("unet", f"{out_name}")
         ckpt_path = folder_paths.get_full_path("checkpoints", ckpt_name)
 
@@ -335,12 +335,9 @@ class CoreMLConverter(COREML_NODE):
             lora_weights=lora_weights,
             attn_impl=attention_implementation,
             config_path=config_path,
+            quantize_nbits=quantize_nbits,
         )
-        unet_target_path = converter.compile_model(
-            out_path=unet_out_path, out_name=out_name, submodule_name="unet"
-        )
-
-        return (CoreMLModel(unet_target_path, compute_unit, "compiled"),)
+        return (CoreMLModel(unet_out_path, compute_unit),)
 
     @staticmethod
     def lora_path(lora_name):

pyproject.toml

@@ -0,0 +1,56 @@
+[project]
+name = "comfyui-coremlsuite"
+description = "This extension contains a set of custom nodes for ComfyUI that allow you to use Core ML models in your ComfyUI workflows."
+version = "2.0.2"
+license = "MIT"
+requires-python = ">=3.12,<3.13"
+packages = [{ include = "coreml_suite" }]
+dependencies = [
+    "torch>=2.7,<2.8",
+    "coremltools>=9,<10",
+    "numpy>=2,<3",
+    "diffusers>=0.30",
+    "peft>=0.13",
+    "omegaconf>=2.3",
+    "transformers>=4.44",
+]
+
+[project.urls]
+Repository = "https://github.com/aszc-dev/ComfyUI-CoreMLSuite"
+
+[tool.comfy]
+PublisherId = "aszc-dev"
+DisplayName = "ComfyUI-CoreMLSuite"
+Icon = ""
+requires-comfyui = ">=0.3.27"
+
+[dependency-groups]
+dev = [
+    "pillow>=12.2.0",
+    "psutil>=7.2.2",
+    "pytest>=9.0.3",
+]
+comfy = [
+    "comfyui-frontend-package==1.14.6",
+    "torchvision",
+    "torchaudio",
+    "torchsde",
+    "einops",
+    "tokenizers>=0.13.3",
+    "safetensors>=0.4.2",
+    "aiohttp>=3.11.8",
+    "yarl>=1.18.0",
+    "kornia>=0.7.1",
+    "spandrel",
+    "soundfile",
+    "sentencepiece",
+]
+
+[tool.pytest.ini_options]
+markers = [
+    "unit: framework-free unit test (Tier 0)",
+    "smoke: macOS-ARM smoke test on a synthetic micro-model (Tier 1)",
+    "m2: requires Apple Silicon + Neural Engine (Tier 2)",
+]
+testpaths = ["tests"]
+addopts = ["--import-mode=importlib", "--confcutdir=tests"]

requirements.txt

@@ -1,6 +1,7 @@
-git+https://github.com/apple/ml-stable-diffusion.git
-coremltools>=7.1
-overrides
-diffusers>=0.22
-peft>=0.6.2
+torch>=2.7,<2.8
+coremltools>=9,<10
+numpy>=2,<3
+diffusers>=0.30
+peft>=0.13
 omegaconf>=2.3
+transformers>=4.44

tests/conftest.py

@@ -0,0 +1,61 @@
+"""Pytest bootstrap for ComfyUI-CoreMLSuite tests.
+
+- Adds the ComfyUI checkout to sys.path so the framework-coupled modules
+  that transitively import `comfy.*` resolve when pytest is invoked from
+  this package's root.
+- Auto-applies tier markers based on the directory a test lives in, so
+  individual files don't have to repeat @pytest.mark.unit / .smoke.
+"""
+import sys
+from pathlib import Path
+
+import pytest
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+COMFY_DIR = REPO_ROOT.parents[1]
+
+for p in (str(COMFY_DIR), str(REPO_ROOT)):
+    if p not in sys.path:
+        sys.path.insert(0, p)
+
+
+_TIER_BY_DIR = {
+    "tests/unit": "unit",
+    "tests/m2": "m2",
+    "tests/integration": "m2",
+    "tests/smoke": "smoke",
+}
+
+# When the user asks for a single tier (-m unit / -m smoke), skip the other
+# directories at collection time. Tier-0 cannot afford to import tests/smoke
+# files because they pull in coremltools which Linux CI won't have.
+_TIER_DIRS = {
+    "unit": ("/tests/unit/",),
+    "m2": ("/tests/m2/", "/tests/integration/"),
+    "smoke": ("/tests/smoke/",),
+}
+
+
+def pytest_ignore_collect(collection_path, config):
+    expr = config.option.markexpr
+    if expr not in _TIER_DIRS:
+        return None
+    allowed = _TIER_DIRS[expr]
+    rel = str(collection_path).replace("\\", "/")
+    if "/tests/" not in rel:
+        return None
+    # Always allow tests/ root + the tier's own dirs.
+    if rel.endswith("/tests"):
+        return None
+    if any(frag in rel + "/" for frag in allowed):
+        return None
+    return True
+
+
+def pytest_collection_modifyitems(config, items):
+    for item in items:
+        path = str(item.fspath).replace("\\", "/")
+        for fragment, marker in _TIER_BY_DIR.items():
+            if f"/{fragment}/" in path:
+                item.add_marker(getattr(pytest.mark, marker))
+                break

tests/integration/test_basic_conversion_1_5.py

@@ -1,72 +0,0 @@
-import json
-import os
-
-import pytest
-import requests
-from PIL import Image
-import numpy as np
-
-from folder_paths import get_save_image_path, get_output_directory
-
-IMAGE_PREFIX = "E2E-1.5-CoreML"
-
-
-class OutputImageRepository:
-    def __init__(self, name_prefix):
-        self.name_prefix = name_prefix
-
-    def list_images(self):
-        full_output_folder, _, _, _, _ = get_save_image_path(
-            self.name_prefix, get_output_directory(), 512, 512
-        )
-        return full_output_folder, os.listdir(full_output_folder)
-
-    def delete_images(self):
-        full_output_folder, images = self.list_images()
-        for image in images:
-            os.remove(os.path.join(full_output_folder, image))
-
-
-@pytest.fixture(scope="module")
-def output_image_repository():
-    repo = OutputImageRepository(IMAGE_PREFIX)
-    yield repo
-    repo.delete_images()
-
-
-def test_basic_conversion_1_5(output_image_repository):
-    with open("integration/workflows/e2e-1.5-basic-conversion.json") as f:
-        prompt = json.load(f)
-    queue_prompt(prompt)
-
-    full_output_folder, images = output_image_repository.list_images()
-    assert len(images) == 2
-    assert all(image.startswith(IMAGE_PREFIX) for image in images)
-    assert all(image.endswith(".png") for image in images)
-    assert all(
-        os.path.isfile(os.path.join(full_output_folder, image)) for image in images
-    )
-
-    image1 = Image.open(os.path.join(full_output_folder, images[0]))
-    image2 = Image.open(os.path.join(full_output_folder, images[1]))
-    assert psnr(np.array(image1), np.array(image2)) > 30
-    assert psnr(np.array(image2), np.array(image1)) > 30
-
-
-def psnr(img1, img2):
-    mse = np.mean((img1 - img2) ** 2)
-    if mse == 0:
-        return 100
-    PIXEL_MAX = 255.0
-    return 20 * np.log10(PIXEL_MAX / np.sqrt(mse))
-
-
-def queue_prompt(prompt: dict):
-    p = {"prompt": prompt}
-    data = json.dumps(p).encode("utf-8")
-    req = requests.post("http://localhost:8188/prompt", data=data)
-    assert req.status_code == 200
-    while True:
-        req = requests.get("http://localhost:8188/prompt")
-        if req.json()["exec_info"]["queue_remaining"] == 0:
-            break

tests/integration/workflows/e2e-1.5-basic-conversion.json

@@ -179,4 +179,4 @@
       "title": "Save Image"
     }
   }
-}
+}

tests/m2/goldens/sd15_seed42.sha256

@@ -0,0 +1 @@
+e89344e544d4edfbd3ebe9a1c78dadb2729f53549666052b74ac7308f326f4fc

tests/m2/test_golden_image.py

@@ -0,0 +1,170 @@
+"""[M2-ANE] golden-image anchor.
+
+Runs the e2e SD1.5 + CoreML workflow against a local ComfyUI server, fetches
+the generated PNG, and asserts both:
+  - byte-identical SHA256 against the stored golden, OR
+  - PSNR >= GOLDEN_PSNR_MIN_DB against the stored golden PNG.
+
+The hash is the strict gate (a refactor that doesn't touch the math
+should hit it). PSNR is the soft gate that tolerates the drift a
+toolchain bump injects through different MIL graphs / kernel selection
+/ fp accumulation order — anything below the threshold is treated as a
+regression.
+
+The 20 dB default absorbs Apple Neural Engine run-to-run nondeterminism:
+the same model and seed can drift several dB between runs as the 20
+sampling steps amplify tiny per-step UNet differences (kernel selection /
+fp accumulation order). Same-scene ANE outputs have been observed at
+~23 dB, so 20 leaves margin while still catching gross regressions — a
+broken image lands far lower. Bump it up for pure-refactor PRs that must
+not change math; down for toolchain bumps.
+
+Skips entirely on non-Apple-Silicon hosts or when the server / converted
+model is missing, so the unit lane on Linux still passes.
+
+The first run with no golden writes one and fails so it's reviewed before
+being committed.
+"""
+import hashlib
+import json
+import os
+import platform
+import shutil
+import time
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+import numpy as np
+import pytest
+from PIL import Image
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+COMFY_DIR = Path(os.environ.get("COMFY_DIR", REPO_ROOT.parents[1])).resolve()
+COMFY_HOST = os.environ.get("COMFY_HOST", "localhost")
+COMFY_PORT = int(os.environ.get("COMFY_PORT", "8188"))
+COMFY_URL = f"http://{COMFY_HOST}:{COMFY_PORT}"
+
+CKPT_NAME = os.environ.get("CKPT_NAME", "v1-5-pruned-emaonly.safetensors")
+WORKFLOW_PATH = (
+    REPO_ROOT / "tests" / "integration" / "workflows" / "e2e-1.5-basic-conversion.json"
+)
+GOLDEN_DIR = Path(__file__).parent / "goldens"
+GOLDEN_HASH_PATH = GOLDEN_DIR / "sd15_seed42.sha256"
+GOLDEN_PNG_PATH = GOLDEN_DIR / "sd15_seed42.png"
+GOLDEN_PSNR_MIN_DB = float(os.environ.get("GOLDEN_PSNR_MIN_DB", "20"))
+SEED = 42
+
+
+def _server_reachable() -> bool:
+    try:
+        with urllib.request.urlopen(f"{COMFY_URL}/prompt", timeout=3) as r:
+            return r.status == 200
+    except (urllib.error.URLError, urllib.error.HTTPError, ConnectionError):
+        return False
+
+
+@pytest.fixture(scope="module")
+def comfy_server():
+    if platform.machine() != "arm64":
+        pytest.skip("requires Apple Silicon")
+    if not _server_reachable():
+        pytest.skip(f"ComfyUI server not reachable at {COMFY_URL}")
+    return COMFY_URL
+
+
+def _http_post_json(path: str, payload: dict) -> dict:
+    data = json.dumps(payload).encode("utf-8")
+    req = urllib.request.Request(
+        f"{COMFY_URL}{path}", data=data,
+        headers={"Content-Type": "application/json"}, method="POST",
+    )
+    with urllib.request.urlopen(req, timeout=300) as r:
+        return json.loads(r.read().decode())
+
+
+def _http_get_json(path: str, timeout: int = 300) -> dict:
+    """ComfyUI runs UNet inference on its single asyncio loop, so GET /prompt
+    blocks while the queued prompt is executing. Use a generous timeout."""
+    with urllib.request.urlopen(f"{COMFY_URL}{path}", timeout=timeout) as r:
+        return json.loads(r.read().decode())
+
+
+def _drain_queue(timeout_s: int = 600) -> None:
+    deadline = time.time() + timeout_s
+    while time.time() < deadline:
+        try:
+            q = _http_get_json("/prompt")
+        except (urllib.error.URLError, TimeoutError):
+            # Transient block while server executes; retry until our overall
+            # deadline expires.
+            continue
+        if q.get("exec_info", {}).get("queue_remaining", -1) == 0:
+            return
+        time.sleep(2)
+    raise TimeoutError(f"queue did not drain within {timeout_s}s")
+
+
+def _post_workflow_and_collect_png() -> bytes:
+    workflow = json.loads(WORKFLOW_PATH.read_text())
+    for nid in ("4", "10"):
+        if nid in workflow:
+            workflow[nid]["inputs"]["ckpt_name"] = CKPT_NAME
+    for nid in ("3", "11"):
+        if nid in workflow and "seed" in workflow[nid].get("inputs", {}):
+            workflow[nid]["inputs"]["seed"] = SEED
+    # Drop the MPS reference branch — only the Core ML pipeline is needed here.
+    for nid in ("3", "8", "9"):
+        workflow.pop(nid, None)
+
+    _http_post_json("/prompt", {"prompt": workflow})
+    _drain_queue()
+
+    comfy_out = COMFY_DIR / "output"
+    matches = sorted(comfy_out.glob("E2E-1.5-CoreML_*.png"), reverse=True)
+    if not matches:
+        raise FileNotFoundError(f"no Core ML image under {comfy_out}")
+    return matches[0].read_bytes()
+
+
+def _psnr(a: np.ndarray, b: np.ndarray) -> float:
+    mse = float(np.mean((a.astype(np.float64) - b.astype(np.float64)) ** 2))
+    if mse == 0:
+        return 100.0
+    return 20.0 * float(np.log10(255.0 / np.sqrt(mse)))
+
+
+def test_sd15_seed42_image_matches_golden(comfy_server):
+    GOLDEN_DIR.mkdir(parents=True, exist_ok=True)
+    png_bytes = _post_workflow_and_collect_png()
+    sha = hashlib.sha256(png_bytes).hexdigest()
+
+    if not GOLDEN_HASH_PATH.exists() or not GOLDEN_PNG_PATH.exists():
+        GOLDEN_HASH_PATH.write_text(sha + "\n")
+        # Persist the PNG too for visual diffing + PSNR.
+        tmp_path = Path(__file__).parent / "_latest_generated.png"
+        tmp_path.write_bytes(png_bytes)
+        shutil.copy2(tmp_path, GOLDEN_PNG_PATH)
+        pytest.fail(
+            f"No golden present; wrote {GOLDEN_HASH_PATH.name} and "
+            f"{GOLDEN_PNG_PATH.name}. Review the image and re-run."
+        )
+
+    expected_hash = GOLDEN_HASH_PATH.read_text().strip()
+    if sha == expected_hash:
+        return
+
+    # Hash drift: fall back to PSNR to distinguish a refactor-safe rounding
+    # change from a real regression.
+    a = np.array(Image.open(GOLDEN_PNG_PATH).convert("RGB"))
+    b_path = Path(__file__).parent / "_latest_generated.png"
+    b_path.write_bytes(png_bytes)
+    b = np.array(Image.open(b_path).convert("RGB"))
+    if a.shape != b.shape:
+        pytest.fail(f"shape mismatch: golden={a.shape} actual={b.shape}")
+    psnr_db = _psnr(a, b)
+    assert psnr_db >= GOLDEN_PSNR_MIN_DB, (
+        f"hash drifted (got {sha[:12]}.., expected {expected_hash[:12]}..) and "
+        f"PSNR {psnr_db:.2f} dB < {GOLDEN_PSNR_MIN_DB} dB threshold; "
+        f"diff PNG at {b_path}"
+    )

tests/smoke/test_split_einsum_attention.py

@@ -0,0 +1,41 @@
+import platform
+
+import pytest
+import torch
+from diffusers.models.attention_processor import Attention, AttnProcessor
+
+from coreml_suite.conversion.attention import (
+    SplitEinsumAttnProcessor,
+    SplitEinsumV2AttnProcessor,
+)
+
+
+pytestmark = pytest.mark.skipif(
+    platform.system() != "Darwin" or platform.machine() != "arm64",
+    reason="Tier 1 requires macOS on Apple Silicon",
+)
+
+
+@pytest.mark.parametrize(
+    "processor",
+    [
+        SplitEinsumAttnProcessor(),
+        SplitEinsumV2AttnProcessor(),
+    ],
+)
+def test_split_einsum_processor_matches_diffusers_attention(processor):
+    torch.manual_seed(0)
+    reference = Attention(query_dim=32, heads=4, dim_head=8, dropout=0.0)
+    reference.set_processor(AttnProcessor())
+
+    candidate = Attention(query_dim=32, heads=4, dim_head=8, dropout=0.0)
+    candidate.load_state_dict(reference.state_dict())
+    candidate.set_processor(processor)
+
+    hidden_states = torch.randn(2, 17, 32)
+    encoder_hidden_states = torch.randn(2, 11, 32)
+
+    expected = reference(hidden_states, encoder_hidden_states=encoder_hidden_states)
+    actual = candidate(hidden_states, encoder_hidden_states=encoder_hidden_states)
+
+    assert torch.allclose(actual, expected, atol=1e-5)

tests/smoke/test_synthetic_unet.py

@@ -0,0 +1,138 @@
+"""Tier 1 smoke: convert a synthetic micro-UNet through coremltools and load
+it back with CoreMLSuite's runtime CoreMLModel wrapper.
+
+Purpose: catch API breakage in coremltools *without* needing a real SD
+checkpoint, the ANE, or a converted .mlmodelc on disk.
+Runs in minutes on a hosted macOS-ARM runner (no Apple internal stuff).
+
+What it asserts:
+  - coremltools.convert still accepts the call shape we use today
+  - the resulting .mlpackage round-trips through CoreMLSuite's CoreMLModel
+  - expected_inputs exposes the input names/shapes we declared
+  - calling the model returns the named output (`noise_pred`)
+
+Auto-skips on non-Apple-Silicon hosts so Tier 0 CI on Linux ignores it.
+"""
+import platform
+import shutil
+from types import SimpleNamespace
+
+import numpy as np
+import pytest
+import torch
+import torch.nn as nn
+
+from coreml_suite.conversion.unet import CoreMLUNetWrapper
+
+
+pytestmark = pytest.mark.skipif(
+    platform.system() != "Darwin" or platform.machine() != "arm64",
+    reason="Tier 1 requires macOS on Apple Silicon",
+)
+
+
+# Tiny shapes — large enough to exercise conv2d + linear + addition kernels in
+# coremltools, small enough that conversion finishes in seconds on CPU.
+SAMPLE_SHAPE = (1, 4, 8, 8)
+TIMESTEP_SHAPE = (1,)
+ENCODER_SHAPE = (1, 4, 64)  # native diffusers encoder_hidden_states (batch, tokens, hidden)
+OUT_NAME = "noise_pred"
+
+
+class TinyUNet(nn.Module):
+    """Minimal UNet-shaped graph: conv -> add(time+context) -> conv.
+
+    Not a real diffusion model. Just enough op variety to exercise the
+    PyTorch -> MIL frontend in coremltools and confirm we can still wire
+    the inputs/outputs the way CoreMLSuite's runtime expects.
+    """
+
+    def __init__(self):
+        super().__init__()
+        self.conv_in = nn.Conv2d(4, 8, kernel_size=3, padding=1)
+        self.conv_out = nn.Conv2d(8, 4, kernel_size=3, padding=1)
+        self.time_proj = nn.Linear(1, 8)
+        self.text_proj = nn.Linear(64, 8)
+
+    def forward(
+        self,
+        sample,
+        timestep,
+        encoder_hidden_states,
+        timestep_cond=None,
+        added_cond_kwargs=None,
+        down_block_additional_residuals=None,
+        mid_block_additional_residual=None,
+        return_dict=True,
+    ):
+        h = self.conv_in(sample)
+        t_emb = self.time_proj(timestep.unsqueeze(-1)).view(1, 8, 1, 1)
+        c_emb = self.text_proj(encoder_hidden_states.mean(1)).view(1, 8, 1, 1)
+        h = h + t_emb + c_emb
+        return (self.conv_out(h),)
+
+
+@pytest.fixture(scope="module")
+def tiny_mlpackage(tmp_path_factory):
+    """Convert TinyUNet once per test session and reuse the .mlpackage."""
+    import coremltools as ct
+
+    torch.manual_seed(0)
+    model = CoreMLUNetWrapper(
+        TinyUNet().eval(),
+        SimpleNamespace(name="SD15"),
+    )
+    example = (
+        torch.randn(*SAMPLE_SHAPE),
+        torch.randn(*TIMESTEP_SHAPE),
+        torch.randn(*ENCODER_SHAPE),
+    )
+    traced = torch.jit.trace(model, example)
+
+    mlmodel = ct.convert(
+        traced,
+        inputs=[
+            ct.TensorType(name="sample", shape=SAMPLE_SHAPE, dtype=np.float16),
+            ct.TensorType(name="timestep", shape=TIMESTEP_SHAPE, dtype=np.float16),
+            ct.TensorType(name="encoder_hidden_states", shape=ENCODER_SHAPE, dtype=np.float16),
+        ],
+        outputs=[ct.TensorType(name=OUT_NAME, dtype=np.float16)],
+        compute_units=ct.ComputeUnit.CPU_ONLY,
+        compute_precision=ct.precision.FLOAT16,
+        convert_to="mlprogram",
+        minimum_deployment_target=ct.target.macOS13,
+    )
+
+    out_dir = tmp_path_factory.mktemp("tiny_unet")
+    pkg_path = out_dir / "tiny.mlpackage"
+    mlmodel.save(str(pkg_path))
+    yield pkg_path
+    shutil.rmtree(out_dir, ignore_errors=True)
+
+
+def test_coremltools_convert_round_trips_via_coreml_model(tiny_mlpackage):
+    from coreml_suite.coreml_model import CoreMLModel
+
+    model = CoreMLModel(str(tiny_mlpackage), "CPU_ONLY")
+
+    # expected_inputs is the contract our wrappers depend on. Lock the shape
+    # of the dict + a sample entry.
+    expected = dict(model.expected_inputs)
+    assert set(expected.keys()) == {"sample", "timestep", "encoder_hidden_states"}
+    assert tuple(expected["sample"]["shape"]) == SAMPLE_SHAPE
+    assert tuple(expected["timestep"]["shape"]) == TIMESTEP_SHAPE
+    assert tuple(expected["encoder_hidden_states"]["shape"]) == ENCODER_SHAPE
+
+    # Forward pass: drive the model the way CoreMLModelWrapper does.
+    rng = np.random.default_rng(0)
+    inputs = {
+        "sample": rng.standard_normal(SAMPLE_SHAPE).astype(np.float16),
+        "timestep": rng.standard_normal(TIMESTEP_SHAPE).astype(np.float16),
+        "encoder_hidden_states": rng.standard_normal(ENCODER_SHAPE).astype(np.float16),
+    }
+    out = model(**inputs)
+    assert isinstance(out, dict), f"unexpected output type: {type(out)}"
+    assert OUT_NAME in out, f"missing output {OUT_NAME!r}; got {sorted(out)}"
+    assert out[OUT_NAME].shape == SAMPLE_SHAPE, (
+        f"output shape drift: got {out[OUT_NAME].shape}, expected {SAMPLE_SHAPE}"
+    )

tests/unit/test_characterization_controlnet.py

@@ -0,0 +1,186 @@
+"""Characterization tests for coreml_suite.controlnet.
+
+Locks shapes + dtypes + zero-fill behavior of expand_inputs / no_control /
+extract_residual_kwargs / chunk_control. These pure helpers feed the Core ML
+UNet's additional_residual_N inputs; any drift here silently breaks
+ControlNet-based workflows.
+"""
+import numpy as np
+import pytest
+import torch
+
+from coreml_suite.core.controlnet import (
+    chunk_control,
+    expand_inputs,
+    extract_residual_kwargs,
+    no_control,
+)
+
+
+@pytest.fixture(autouse=True)
+def _deterministic_seed():
+    torch.manual_seed(0)
+    np.random.seed(0)
+
+
+SD15_RESIDUAL_SPEC = {
+    "additional_residual_0": {"shape": (2, 320, 64, 64)},
+    "additional_residual_1": {"shape": (2, 640, 32, 32)},
+    "additional_residual_2": {"shape": (2, 1280, 8, 8)},
+}
+NON_RESIDUAL_SPEC = {
+    "sample": {"shape": (2, 4, 64, 64)},
+    "encoder_hidden_states": {"shape": (2, 77, 768)},
+}
+
+
+# ---------- expand_inputs ----------------------------------------------------
+
+
+def test_expand_inputs_doubles_singleton_numpy():
+    inputs = {"a": np.ones((1, 4), dtype=np.float32)}
+    out = expand_inputs(inputs)
+    assert out["a"].shape == (2, 4)
+    assert np.array_equal(out["a"], np.ones((2, 4)))
+
+
+def test_expand_inputs_doubles_singleton_torch():
+    inputs = {"a": torch.ones(1, 4)}
+    out = expand_inputs(inputs)
+    assert out["a"].shape == (2, 4)
+    assert torch.equal(out["a"], torch.ones(2, 4))
+
+
+def test_expand_inputs_doubles_singleton_list():
+    inputs = {"a": [42]}
+    out = expand_inputs(inputs)
+    assert out["a"] == [42, 42]
+
+
+def test_expand_inputs_skips_already_batched():
+    """batch > 1 inputs are returned unchanged (same object identity)."""
+    arr = np.ones((2, 4), dtype=np.float32)
+    tensor = torch.ones(3, 4)
+    lst = [1, 2]
+    out = expand_inputs({"a": arr, "b": tensor, "c": lst})
+    assert out["a"] is arr
+    assert out["b"] is tensor
+    assert out["c"] is lst
+
+
+def test_expand_inputs_preserves_unknown_value_types():
+    # Strings/None pass through untouched — locks current permissive contract.
+    inputs = {"s": "hello", "none": None, "int": 7}
+    out = expand_inputs(inputs)
+    assert out == {"s": "hello", "none": None, "int": 7}
+
+
+# ---------- no_control -------------------------------------------------------
+
+
+def test_no_control_returns_zero_fp16_for_residuals():
+    out = no_control({**SD15_RESIDUAL_SPEC, **NON_RESIDUAL_SPEC})
+    # Only additional_residual_* keys are produced.
+    assert set(out.keys()) == set(SD15_RESIDUAL_SPEC.keys())
+    for key, spec in SD15_RESIDUAL_SPEC.items():
+        arr = out[key]
+        assert arr.shape == spec["shape"]
+        assert arr.dtype == np.float16
+        assert np.all(arr == 0)
+
+
+def test_no_control_returns_empty_when_no_residuals():
+    out = no_control(NON_RESIDUAL_SPEC)
+    assert out == {}
+
+
+# ---------- extract_residual_kwargs -----------------------------------------
+
+
+def test_extract_residual_kwargs_empty_when_model_has_no_residual_inputs():
+    out = extract_residual_kwargs(NON_RESIDUAL_SPEC, control={"output": [], "middle": []})
+    assert out == {}
+
+
+def test_extract_residual_kwargs_none_control_returns_no_control_shapes():
+    out = extract_residual_kwargs(SD15_RESIDUAL_SPEC, control=None)
+    assert set(out.keys()) == set(SD15_RESIDUAL_SPEC.keys())
+    for key, spec in SD15_RESIDUAL_SPEC.items():
+        assert out[key].shape == spec["shape"]
+        assert out[key].dtype == np.float16
+        assert np.all(out[key] == 0)
+
+
+def test_extract_residual_kwargs_flattens_output_then_middle_and_casts_fp16():
+    """output residuals come first (indexed 0..N-1), then middle residuals
+    (indexed N..M-1). Values come out of CPU as fp16 numpy arrays."""
+    control = {
+        "output": [torch.ones(2, 320, 64, 64) * 0.5, torch.ones(2, 640, 32, 32) * 2.0],
+        "middle": [torch.ones(2, 1280, 8, 8) * -1.0],
+    }
+    out = extract_residual_kwargs(SD15_RESIDUAL_SPEC, control)
+    assert set(out.keys()) == {"additional_residual_0", "additional_residual_1", "additional_residual_2"}
+    assert out["additional_residual_0"].shape == (2, 320, 64, 64)
+    assert out["additional_residual_1"].shape == (2, 640, 32, 32)
+    assert out["additional_residual_2"].shape == (2, 1280, 8, 8)
+    for arr in out.values():
+        assert arr.dtype == np.float16
+    # Locked order: index 0 == first output residual (0.5), index 2 == middle (-1.0).
+    assert np.allclose(out["additional_residual_0"], 0.5)
+    assert np.allclose(out["additional_residual_1"], 2.0)
+    assert np.allclose(out["additional_residual_2"], -1.0)
+
+
+# ---------- chunk_control ----------------------------------------------------
+
+
+def test_chunk_control_none_returns_list_of_nones_with_length_target():
+    """`no_control` path: when there's no control, you get [None] * target_size
+    (NOT [None, None] regardless of target — this is the contract today)."""
+    assert chunk_control(None, 1) == [None]
+    assert chunk_control(None, 2) == [None, None]
+    assert chunk_control(None, 4) == [None, None, None, None]
+
+
+@pytest.mark.parametrize(
+    "batch,target,expected_chunks",
+    [(1, 2, 1), (2, 2, 1), (3, 2, 2), (4, 2, 2), (5, 3, 2), (9, 4, 3)],
+)
+def test_chunk_control_shapes_after_chunking(batch, target, expected_chunks):
+    cn = {
+        "output": [
+            torch.randn(batch, 320, 64, 64),
+            torch.randn(batch, 640, 32, 32),
+        ],
+        "middle": [torch.randn(batch, 1280, 8, 8)],
+    }
+    chunks = chunk_control(cn, target)
+    assert len(chunks) == expected_chunks
+    for c in chunks:
+        assert c["output"][0].shape == (target, 320, 64, 64)
+        assert c["output"][1].shape == (target, 640, 32, 32)
+        assert c["middle"][0].shape == (target, 1280, 8, 8)
+
+
+def test_chunk_control_preserves_keys_order():
+    """Output dicts contain exactly {"output", "middle"} in that order."""
+    cn = {
+        "output": [torch.zeros(2, 4, 4, 4)],
+        "middle": [torch.zeros(2, 4, 4, 4)],
+    }
+    chunks = chunk_control(cn, 2)
+    assert list(chunks[0].keys()) == ["output", "middle"]
+
+
+def test_chunk_control_zero_pads_remainder():
+    """A batch=3, target=2 split puts the third row alongside a zero row."""
+    cn = {
+        "output": [torch.arange(3 * 4).reshape(3, 1, 2, 2).float()],
+        "middle": [torch.arange(3 * 4).reshape(3, 1, 2, 2).float()],
+    }
+    chunks = chunk_control(cn, 2)
+    assert len(chunks) == 2
+    last_out = chunks[-1]["output"][0]
+    # First row is the original third row; second row is padding zeros.
+    assert torch.equal(last_out[0], cn["output"][0][2])
+    assert torch.equal(last_out[1], torch.zeros(1, 2, 2))

tests/unit/test_characterization_inputs.py

@@ -0,0 +1,228 @@
+"""Characterization tests for coreml_suite.models.CoreMLInputs.
+
+Locks the shape transforms applied by chunks() and coreml_kwargs() for the
+four model variants the suite supports: SD1.5, LCM (SD1.5 + timestep_cond),
+SDXL base (time_ids len 6), and SDXL refiner (time_ids len 5).
+
+These contracts feed the Core ML UNet at runtime; if a refactor silently
+re-shapes them, generation breaks.
+"""
+import numpy as np
+import pytest
+import torch
+
+from coreml_suite.core.inputs import CoreMLInputs
+
+
+@pytest.fixture(autouse=True)
+def _deterministic_seed():
+    torch.manual_seed(0)
+    np.random.seed(0)
+
+
+# ---------- expected_inputs fixtures (mirror real model expectations) -------
+
+SD15_EXPECTED = {
+    "sample": {"shape": (2, 4, 64, 64)},
+    "timestep": {"shape": (2,)},
+    "encoder_hidden_states": {"shape": (2, 77, 768)},
+}
+
+SD15_WITH_CN = {
+    **SD15_EXPECTED,
+    "additional_residual_0": {"shape": (2, 320, 64, 64)},
+    "additional_residual_1": {"shape": (2, 640, 32, 32)},
+}
+
+LCM_EXPECTED = {
+    **SD15_EXPECTED,
+    "timestep_cond": {"shape": (2, 256)},
+}
+
+SDXL_BASE_EXPECTED = {
+    "sample": {"shape": (2, 4, 128, 128)},
+    "timestep": {"shape": (2,)},
+    "encoder_hidden_states": {"shape": (2, 77, 2048)},
+    "time_ids": {"shape": (2, 6)},
+    "text_embeds": {"shape": (2, 1280)},
+}
+
+SDXL_REFINER_EXPECTED = {
+    "sample": {"shape": (2, 4, 128, 128)},
+    "timestep": {"shape": (2,)},
+    "encoder_hidden_states": {"shape": (2, 77, 1280)},
+    "time_ids": {"shape": (2, 5)},
+    "text_embeds": {"shape": (2, 1280)},
+}
+
+
+def _sd15_inputs(batch=1, with_control=False, with_ts_cond=False):
+    x = torch.randn(batch, 4, 64, 64)
+    t = torch.full((batch,), 999.0)
+    context = torch.randn(batch, 77, 768)
+    control = None
+    if with_control:
+        control = {
+            "output": [torch.randn(batch, 320, 64, 64), torch.randn(batch, 640, 32, 32)],
+            "middle": [],
+        }
+    kwargs = {}
+    if with_ts_cond:
+        kwargs["timestep_cond"] = torch.randn(batch, 256)
+    return CoreMLInputs(x, t, context, control, **kwargs)
+
+
+def _sdxl_inputs(batch=1, refiner=False):
+    x = torch.randn(batch, 4, 128, 128)
+    t = torch.full((batch,), 999.0)
+    ctx_dim = 1280 if refiner else 2048
+    context = torch.randn(batch, 77, ctx_dim)
+    time_ids_dim = 5 if refiner else 6
+    time_ids = torch.randn(batch, time_ids_dim)
+    text_embeds = torch.randn(batch, 1280)
+    return CoreMLInputs(
+        x, t, context, control=None, time_ids=time_ids, text_embeds=text_embeds
+    )
+
+
+# ---------- coreml_kwargs ---------------------------------------------------
+
+
+def test_coreml_kwargs_sd15_shapes_and_fp16():
+    out = _sd15_inputs(batch=1).coreml_kwargs(SD15_EXPECTED)
+    assert set(out.keys()) == {"sample", "encoder_hidden_states", "timestep"}
+    assert out["sample"].shape == (1, 4, 64, 64)
+    assert out["sample"].dtype == np.float16
+    # encoder_hidden_states keeps Comfy's native (b, seq, dim) layout.
+    assert out["encoder_hidden_states"].shape == (1, 77, 768)
+    assert out["encoder_hidden_states"].dtype == np.float16
+    assert out["timestep"].shape == (1,)
+    assert out["timestep"].dtype == np.float16
+
+
+def test_coreml_kwargs_sd15_with_controlnet_emits_residuals():
+    inputs = _sd15_inputs(batch=1, with_control=True)
+    out = inputs.coreml_kwargs(SD15_WITH_CN)
+    assert "additional_residual_0" in out
+    assert "additional_residual_1" in out
+    assert out["additional_residual_0"].shape == (1, 320, 64, 64)
+    assert out["additional_residual_1"].shape == (1, 640, 32, 32)
+
+
+def test_coreml_kwargs_sd15_without_controlnet_zero_fills_residuals():
+    inputs = _sd15_inputs(batch=1, with_control=False)
+    out = inputs.coreml_kwargs(SD15_WITH_CN)
+    assert np.all(out["additional_residual_0"] == 0)
+    assert np.all(out["additional_residual_1"] == 0)
+
+
+def test_coreml_kwargs_lcm_adds_timestep_cond():
+    inputs = _sd15_inputs(batch=1, with_ts_cond=True)
+    out = inputs.coreml_kwargs(LCM_EXPECTED)
+    assert "timestep_cond" in out
+    assert out["timestep_cond"].shape == (1, 256)
+    assert out["timestep_cond"].dtype == np.float16
+
+
+def test_coreml_kwargs_lcm_skips_timestep_cond_when_not_provided():
+    """timestep_cond is only forwarded when the input supplied one — even if
+    the model's expected_inputs lists it."""
+    inputs = _sd15_inputs(batch=1, with_ts_cond=False)
+    out = inputs.coreml_kwargs(LCM_EXPECTED)
+    assert "timestep_cond" not in out
+
+
+def test_coreml_kwargs_sdxl_base_emits_time_ids_and_text_embeds():
+    out = _sdxl_inputs(batch=1, refiner=False).coreml_kwargs(SDXL_BASE_EXPECTED)
+    assert out["time_ids"].shape == (1, 6)
+    assert out["text_embeds"].shape == (1, 1280)
+    assert out["time_ids"].dtype == np.float16
+    assert out["text_embeds"].dtype == np.float16
+
+
+def test_coreml_kwargs_sdxl_refiner_uses_len5_time_ids():
+    out = _sdxl_inputs(batch=1, refiner=True).coreml_kwargs(SDXL_REFINER_EXPECTED)
+    assert out["time_ids"].shape == (1, 5)
+
+
+# ---------- chunks ----------------------------------------------------------
+
+
+def test_chunks_sd15_pad_to_batch2_returns_one_chunk():
+    chunked = _sd15_inputs(batch=1).chunks(SD15_EXPECTED)
+    assert len(chunked) == 1
+    c = chunked[0]
+    assert c.x.shape == (2, 4, 64, 64)
+    assert c.t.shape == (2,)
+    # context shape: (b, seq, dim) padded along batch dim.
+    assert c.context.shape == (2, 77, 768)
+    assert c.control is None
+    assert c.ts_cond is None
+    assert c.time_ids is None
+    assert c.text_embeds is None
+
+
+def test_chunks_sd15_with_controlnet_chunks_residuals_too():
+    chunked = _sd15_inputs(batch=1, with_control=True).chunks(SD15_EXPECTED)
+    assert len(chunked) == 1
+    cn = chunked[0].control
+    assert cn is not None
+    assert cn["output"][0].shape == (2, 320, 64, 64)
+    assert cn["output"][1].shape == (2, 640, 32, 32)
+
+
+def test_chunks_lcm_carries_timestep_cond_per_chunk():
+    chunked = _sd15_inputs(batch=1, with_ts_cond=True).chunks(LCM_EXPECTED)
+    assert len(chunked) == 1
+    assert chunked[0].ts_cond is not None
+    assert chunked[0].ts_cond.shape == (2, 256)
+
+
+def test_chunks_sdxl_base_propagates_time_ids_and_text_embeds():
+    chunked = _sdxl_inputs(batch=1, refiner=False).chunks(SDXL_BASE_EXPECTED)
+    assert len(chunked) == 1
+    c = chunked[0]
+    assert c.time_ids is not None and c.time_ids.shape == (2, 6)
+    assert c.text_embeds is not None and c.text_embeds.shape == (2, 1280)
+
+
+def test_chunks_sdxl_refiner_uses_len5_time_ids():
+    chunked = _sdxl_inputs(batch=1, refiner=True).chunks(SDXL_REFINER_EXPECTED)
+    assert chunked[0].time_ids.shape == (2, 5)
+
+
+def test_chunks_sdxl_synthesizes_zero_time_ids_when_caller_omits():
+    """If the model expects time_ids but caller passed nothing, the suite
+    fabricates a zero-filled tensor. Lock that fallback."""
+    x = torch.randn(1, 4, 128, 128)
+    t = torch.full((1,), 999.0)
+    context = torch.randn(1, 77, 2048)
+    inputs = CoreMLInputs(x, t, context, control=None)
+    chunked = inputs.chunks(SDXL_BASE_EXPECTED)
+    assert chunked[0].time_ids.shape == (2, 6)
+    assert torch.equal(chunked[0].time_ids, torch.zeros(2, 6))
+    assert chunked[0].text_embeds.shape == (2, 1280)
+    assert torch.equal(chunked[0].text_embeds, torch.zeros(2, 1280))
+
+
+def test_chunks_splits_batch_into_multiple_target2_chunks():
+    """batch=5 with target_batch=2 -> 3 chunks (last padded)."""
+    chunked = _sd15_inputs(batch=5).chunks(SD15_EXPECTED)
+    assert len(chunked) == 3
+    for c in chunked:
+        assert c.x.shape == (2, 4, 64, 64)
+        assert c.context.shape == (2, 77, 768)
+    # Last chunk's second batch row is the zero-pad.
+    assert torch.equal(chunked[-1].x[1], torch.zeros(4, 64, 64))
+
+
+def test_chunks_timestep_is_broadcast_from_first_value():
+    """t is rebuilt from t[0] across all chunks: locks current behavior that
+    discards any per-row timestep variation."""
+    x = torch.randn(2, 4, 64, 64)
+    t = torch.tensor([42.0, 99.0])  # the second value will be lost
+    context = torch.randn(2, 77, 768)
+    inputs = CoreMLInputs(x, t, context, control=None)
+    chunked = inputs.chunks(SD15_EXPECTED)
+    assert chunked[0].t.shape == (2,)
+    assert torch.equal(chunked[0].t, torch.full((2,), 42.0))

tests/unit/test_characterization_latents.py

@@ -0,0 +1,118 @@
+"""Characterization tests for coreml_suite.latents.
+
+Locks the *current* behavior of chunk_batch / merge_chunks — including the
+zero-pad regions and the truncation in merge — so a refactor
+cannot silently shift either contract.
+"""
+import pytest
+import torch
+
+from coreml_suite.core.latents import chunk_batch, merge_chunks
+
+
+@pytest.fixture(autouse=True)
+def _deterministic_seed():
+    torch.manual_seed(0)
+
+
+def _const_tensor(batch, *rest):
+    return torch.arange(batch * 4 * 8 * 8, dtype=torch.float32).reshape(batch, 4, 8, 8)
+
+
+# ---------- chunk_batch ------------------------------------------------------
+
+
+def test_chunk_batch_passthrough_when_shape_matches():
+    x = _const_tensor(2)
+    out = chunk_batch(x, (2, 4, 8, 8))
+    assert len(out) == 1
+    # passthrough: the same object identity is returned (no copy).
+    assert out[0] is x
+
+
+def test_chunk_batch_pads_single_chunk_when_input_smaller():
+    """batch=1, target=2 -> one padded chunk; the second row is exact zero."""
+    x = _const_tensor(1)
+    out = chunk_batch(x, (2, 4, 8, 8))
+    assert len(out) == 1
+    assert out[0].shape == (2, 4, 8, 8)
+    assert torch.equal(out[0][0], x[0])
+    assert torch.equal(out[0][1], torch.zeros(4, 8, 8))
+
+
+def test_chunk_batch_splits_exact_multiple():
+    """batch=4, target=2 -> two chunks, no padding."""
+    x = _const_tensor(4)
+    out = chunk_batch(x, (2, 4, 8, 8))
+    assert len(out) == 2
+    assert out[0].shape == (2, 4, 8, 8)
+    assert out[1].shape == (2, 4, 8, 8)
+    assert torch.equal(out[0], x[:2])
+    assert torch.equal(out[1], x[2:])
+
+
+def test_chunk_batch_pads_remainder_chunk():
+    """batch=5, target=2 -> chunks=[x[0:2], x[2:4]] then [x[4], 0]."""
+    x = _const_tensor(5)
+    out = chunk_batch(x, (2, 4, 8, 8))
+    assert len(out) == 3
+    assert torch.equal(out[0], x[0:2])
+    assert torch.equal(out[1], x[2:4])
+    last = out[-1]
+    assert last.shape == (2, 4, 8, 8)
+    assert torch.equal(last[0], x[4])
+    # The remainder row is zero-padded; lock that exact contract.
+    assert torch.equal(last[1], torch.zeros(4, 8, 8))
+    assert last[1].sum() == 0
+
+
+@pytest.mark.parametrize(
+    "batch_size,target,expected_chunks",
+    [
+        (1, 4, 1),
+        (3, 2, 2),
+        (5, 3, 2),
+        (9, 4, 3),
+    ],
+)
+def test_chunk_batch_pad_region_is_zero(batch_size, target, expected_chunks):
+    x = _const_tensor(batch_size)
+    out = chunk_batch(x, (target, 4, 8, 8))
+    assert len(out) == expected_chunks
+    mod = batch_size % target
+    if mod == 0 and batch_size >= target:
+        return
+    last = out[-1]
+    pad_rows = target - (mod if (mod != 0 and batch_size >= target) else batch_size)
+    pad_region = last[-pad_rows:]
+    assert torch.equal(pad_region, torch.zeros_like(pad_region))
+
+
+# ---------- merge_chunks -----------------------------------------------------
+
+
+def test_merge_chunks_exact_concat():
+    x = _const_tensor(4)
+    chunks = chunk_batch(x, (2, 4, 8, 8))
+    merged = merge_chunks(chunks, x.shape)
+    assert merged.shape == x.shape
+    assert torch.equal(merged, x)
+
+
+def test_merge_chunks_truncates_padding():
+    """Round-trip with a padded last chunk drops the pad rows."""
+    x = _const_tensor(5)
+    chunks = chunk_batch(x, (2, 4, 8, 8))
+    merged = merge_chunks(chunks, x.shape)
+    assert merged.shape == x.shape
+    assert torch.equal(merged, x)
+
+
+def test_merge_chunks_singleton_returns_equal_copy_when_shape_matches():
+    """A singleton chunk list still goes through torch.cat, so we get a new
+    tensor equal to the input — locked here because a refactor might be tempted
+    to short-circuit and accidentally return the same object."""
+    x = _const_tensor(2)
+    out = merge_chunks([x], x.shape)
+    assert torch.equal(out, x)
+    assert out is not x

tests/unit/test_characterization_out_name.py

@@ -0,0 +1,197 @@
+"""Characterization tests for the .mlpackage filename composition.
+
+The filename composition is the pure
+coreml_suite.core.naming.compose_out_name function. CoreMLConverter.convert
+calls it; testing the pure function avoids monkey-patching heavy converter
+internals just to capture the string.
+"""
+import pytest
+
+from coreml_suite.core.naming import compose_out_name, lora_names_from_params
+
+
+# ---------- attention suffixes ----------------------------------------------
+
+
+@pytest.mark.parametrize(
+    "attn_name,suffix",
+    [
+        ("SPLIT_EINSUM", "se"),
+        ("SPLIT_EINSUM_V2", "se2"),
+        ("ORIGINAL", "orig"),
+    ],
+)
+def test_attention_suffix(attn_name, suffix):
+    out = compose_out_name(
+        ckpt_name="dreamshaper_8.safetensors",
+        batch_size=1, width=512, height=512,
+        controlnet_support=False,
+        attention_implementation=attn_name,
+    )
+    assert out == f"dreamshaper_8_1x512x512_{suffix}"
+
+
+# ---------- batch / size ----------------------------------------------------
+
+
+def test_includes_batch_and_size():
+    out = compose_out_name(
+        ckpt_name="dreamshaper_8.safetensors",
+        batch_size=4, width=768, height=1024,
+        controlnet_support=False,
+        attention_implementation="SPLIT_EINSUM",
+    )
+    assert out == "dreamshaper_8_4x768x1024_se"
+
+
+# ---------- ControlNet ------------------------------------------------------
+
+
+def test_appends_cn_suffix_when_controlnet_support_true():
+    out = compose_out_name(
+        ckpt_name="dreamshaper_8.safetensors",
+        batch_size=1, width=512, height=512,
+        controlnet_support=True,
+        attention_implementation="SPLIT_EINSUM",
+    )
+    assert out == "dreamshaper_8_1x512x512_cn_se"
+
+
+# ---------- ckpt name massage -----------------------------------------------
+
+
+def test_drops_extension_at_first_period():
+    out = compose_out_name(
+        ckpt_name="my.checkpoint.v2.safetensors",
+        batch_size=1, width=512, height=512,
+        controlnet_support=False,
+        attention_implementation="SPLIT_EINSUM",
+    )
+    assert out == "my_1x512x512_se"
+
+
+def test_replaces_spaces_with_underscores():
+    out = compose_out_name(
+        ckpt_name="dream shaper 8.safetensors",
+        batch_size=1, width=512, height=512,
+        controlnet_support=False,
+        attention_implementation="SPLIT_EINSUM",
+    )
+    assert out == "dream_shaper_8_1x512x512_se"
+
+
+# ---------- LoRA suffixes ---------------------------------------------------
+
+
+def test_single_lora():
+    out = compose_out_name(
+        ckpt_name="dreamshaper_8.safetensors",
+        batch_size=1, width=512, height=512,
+        controlnet_support=False,
+        attention_implementation="SPLIT_EINSUM",
+        lora_names=["epi_noiseoffset.safetensors"],
+    )
+    assert out == "dreamshaper_8_epi_noiseoffset_1x512x512_se"
+
+
+def test_multiple_loras_sorted():
+    out = compose_out_name(
+        ckpt_name="dreamshaper_8.safetensors",
+        batch_size=1, width=512, height=512,
+        controlnet_support=False,
+        attention_implementation="SPLIT_EINSUM",
+        lora_names=["zoom.safetensors", "alpha.safetensors", "moody.safetensors"],
+    )
+    assert out == "dreamshaper_8_alpha_moody_zoom_1x512x512_se"
+
+
+def test_lora_plus_controlnet():
+    out = compose_out_name(
+        ckpt_name="dreamshaper_8.safetensors",
+        batch_size=1, width=512, height=512,
+        controlnet_support=True,
+        attention_implementation="SPLIT_EINSUM",
+        lora_names=["a.safetensors"],
+    )
+    assert out == "dreamshaper_8_a_1x512x512_cn_se"
+
+
+# ---------- sdxl combinations -----------------------------------------------
+
+
+def test_sdxl_1024_original_gpu():
+    out = compose_out_name(
+        ckpt_name="sd_xl_base_1.0.safetensors",
+        batch_size=1, width=1024, height=1024,
+        controlnet_support=False,
+        attention_implementation="ORIGINAL",
+    )
+    assert out == "sd_xl_base_1_1x1024x1024_orig"
+
+
+# ---------- lora_names_from_params helper ----------------------------------
+
+
+def test_lora_names_from_params_sorts_by_name():
+    names = lora_names_from_params([
+        ("zebra.safetensors", 1.0),
+        ("apple.safetensors", 0.5),
+        ("mango.safetensors", 0.7),
+    ])
+    assert names == ["apple.safetensors", "mango.safetensors", "zebra.safetensors"]
+
+
+def test_lora_names_from_params_empty_list():
+    assert lora_names_from_params([]) == []
+
+
+# ---------- quantize_nbits suffix ------------------------------------------
+
+
+def test_quantize_nbits_none_appends_nothing():
+    """'none' is the default and must keep the unquantized filename so
+    existing cached .mlpackages still resolve."""
+    out = compose_out_name(
+        ckpt_name="dreamshaper_8.safetensors",
+        batch_size=1, width=512, height=512,
+        controlnet_support=False,
+        attention_implementation="SPLIT_EINSUM",
+        quantize_nbits="none",
+    )
+    assert out == "dreamshaper_8_1x512x512_se"
+
+
+@pytest.mark.parametrize("nbits,suffix", [("4", "_q4"), ("6", "_q6"), ("8", "_q8")])
+def test_quantize_nbits_appends_q_suffix(nbits, suffix):
+    out = compose_out_name(
+        ckpt_name="dreamshaper_8.safetensors",
+        batch_size=1, width=512, height=512,
+        controlnet_support=False,
+        attention_implementation="SPLIT_EINSUM",
+        quantize_nbits=nbits,
+    )
+    assert out == f"dreamshaper_8_1x512x512_se{suffix}"
+
+
+def test_quantize_nbits_with_controlnet_and_lora():
+    out = compose_out_name(
+        ckpt_name="dreamshaper_8.safetensors",
+        batch_size=1, width=512, height=512,
+        controlnet_support=True,
+        attention_implementation="SPLIT_EINSUM",
+        lora_names=["a.safetensors"],
+        quantize_nbits="6",
+    )
+    assert out == "dreamshaper_8_a_1x512x512_cn_se_q6"
+
+
+def test_quantize_nbits_invalid_raises():
+    import pytest as _pytest
+    with _pytest.raises(ValueError, match="quantize_nbits"):
+        compose_out_name(
+            ckpt_name="x.safetensors",
+            batch_size=1, width=512, height=512,
+            controlnet_support=False,
+            attention_implementation="SPLIT_EINSUM",
+            quantize_nbits="16",  # not in {none, 8, 6, 4}
+        )

tests/unit/test_characterization_sdxl_options.py

@@ -0,0 +1,127 @@
+"""Characterization tests for the SDXL options math.
+
+The SDXL time_ids / text_embeds math lives in
+coreml_suite.core.sdxl as pure builders. The framework adapter
+add_sdxl_model_options lives in models.py; here we just lock the pure
+math.
+"""
+import inspect
+
+import pytest
+import torch
+
+from coreml_suite.core.sdxl import (
+    build_sdxl_text_embeds,
+    build_sdxl_time_ids,
+    sdxl_model_function_wrapper,
+)
+
+
+@pytest.fixture(autouse=True)
+def _deterministic_seed():
+    torch.manual_seed(0)
+
+
+# ---------- build_sdxl_time_ids: base (len 6) -------------------------------
+
+
+def test_build_time_ids_base_defaults():
+    out = build_sdxl_time_ids({}, {}, is_base=True, is_refiner=False)
+    expected = torch.tensor([[768, 768, 0, 0, 768, 768], [768, 768, 0, 0, 768, 768]])
+    assert out.shape == (2, 6)
+    assert torch.equal(out, expected)
+
+
+def test_build_time_ids_base_respects_overrides():
+    pos = {"height": 1024, "width": 512, "crop_h": 8, "crop_w": 4,
+           "target_height": 1024, "target_width": 1024}
+    neg = {"height": 256, "width": 256, "crop_h": 0, "crop_w": 0,
+           "target_height": 256, "target_width": 256}
+    out = build_sdxl_time_ids(pos, neg, is_base=True, is_refiner=False)
+    expected = torch.tensor([[1024, 512, 8, 4, 1024, 1024], [256, 256, 0, 0, 256, 256]])
+    assert torch.equal(out, expected)
+
+
+# ---------- build_sdxl_time_ids: refiner (len 5) ----------------------------
+
+
+def test_build_time_ids_refiner_defaults():
+    out = build_sdxl_time_ids({}, {}, is_base=False, is_refiner=True)
+    expected = torch.tensor([[768, 768, 0, 0, 6.0], [768, 768, 0, 0, 2.5]])
+    assert out.shape == (2, 5)
+    assert torch.equal(out, expected)
+
+
+def test_build_time_ids_refiner_respects_aesthetic_score():
+    pos = {"aesthetic_score": 8.5}
+    neg = {"aesthetic_score": 1.5}
+    out = build_sdxl_time_ids(pos, neg, is_base=False, is_refiner=True)
+    expected = torch.tensor([[768, 768, 0, 0, 8.5], [768, 768, 0, 0, 1.5]])
+    assert torch.equal(out, expected)
+
+
+# ---------- build_sdxl_time_ids: edge case ----------------------------------
+
+
+def test_build_time_ids_neither_base_nor_refiner_returns_len4():
+    out = build_sdxl_time_ids({}, {}, is_base=False, is_refiner=False)
+    assert out.shape == (2, 4)
+
+
+# ---------- build_sdxl_text_embeds ------------------------------------------
+
+
+def test_text_embeds_concat_pos_then_neg():
+    pos = torch.full((1, 1280), 1.0)
+    neg = torch.full((1, 1280), -1.0)
+    out = build_sdxl_text_embeds(pos, neg)
+    assert out.shape == (2, 1280)
+    assert torch.equal(out[0], pos[0])
+    assert torch.equal(out[1], neg[0])
+
+
+# ---------- sdxl_model_function_wrapper closure -----------------------------
+
+
+def test_wrapper_captures_time_ids_text_embeds_refiner_via_closure():
+    time_ids = torch.zeros(2, 6)
+    text_embeds = torch.zeros(2, 1280)
+    wrapper = sdxl_model_function_wrapper(time_ids, text_embeds, refiner=False)
+    closure = inspect.getclosurevars(wrapper).nonlocals
+    assert closure["time_ids"] is time_ids
+    assert closure["text_embeds"] is text_embeds
+    assert closure["refiner"] is False
+
+
+def test_wrapper_returns_zero_when_context_missing():
+    """When c_crossattn is None the wrapper short-circuits to zeros_like(x).
+    Locked here because the refactor mustn't change this default."""
+    wrapper = sdxl_model_function_wrapper(torch.zeros(2, 6), torch.zeros(2, 1280))
+    x = torch.randn(2, 4, 16, 16)
+    out = wrapper(
+        model_function=lambda *a, **kw: pytest.fail("model_function must not run"),
+        params={"input": x, "timestep": torch.zeros(2), "c": {}},
+    )
+    assert torch.equal(out, torch.zeros_like(x))
+
+
+def test_wrapper_refiner_truncates_context_to_g_clip():
+    """refiner=True slices c_crossattn[:, :, 768:] before forwarding."""
+    captured = {}
+
+    def fake_model(x, t, **c):
+        captured["context_shape"] = c["c_crossattn"].shape
+        captured["time_ids_shape"] = c["time_ids"].shape
+        return x
+
+    wrapper = sdxl_model_function_wrapper(
+        torch.zeros(2, 5), torch.zeros(2, 1280), refiner=True
+    )
+    x = torch.randn(2, 4, 16, 16)
+    context = torch.randn(2, 77, 2048)  # 768 + 1280 dims
+    wrapper(
+        model_function=fake_model,
+        params={"input": x, "timestep": torch.zeros(2), "c": {"c_crossattn": context}},
+    )
+    assert captured["context_shape"] == (2, 77, 1280)
+    assert captured["time_ids_shape"] == (2, 5)

tests/unit/test_chunks.py

@@ -1,37 +1,34 @@
-import pytest
+"""Smoke tests for the pure batch-chunking helpers in coreml_suite.core.
 
+Uses torch.device('cpu') instead of comfy.model_management.get_torch_device
+so Tier 0 runs without ComfyUI.
+"""
+import pytest
 import torch
 
-from comfy.model_management import get_torch_device
-from coreml_suite.latents import chunk_batch, merge_chunks
-from coreml_suite.controlnet import chunk_control
-from coreml_suite.models import (
-    CoreMLInputs,
-)
-from coreml_suite.config import get_model_config
+from coreml_suite.core.controlnet import chunk_control
+from coreml_suite.core.inputs import CoreMLInputs
+from coreml_suite.core.latents import chunk_batch, merge_chunks
+
+
+CPU = torch.device("cpu")
 
 
 @pytest.fixture
 def expected_inputs():
-    expected = {
+    return {
         "sample": {"shape": (2, 4, 64, 64)},
         "timestep": {"shape": (2,)},
         "timestep_cond": {"shape": (2, 256)},
-        "encoder_hidden_states": {"shape": (2, 768, 1, 77)},
+        "encoder_hidden_states": {"shape": (2, 77, 768)},
         "additional_residual_0": {"shape": (2, 320, 64, 64)},
         "additional_residual_1": {"shape": (2, 640, 32, 32)},
     }
-    return expected
-
-
-@pytest.fixture
-def model_config():
-    return get_model_config()
 
 
 @pytest.mark.parametrize("batch_size", [1, 2, 4, 5, 9])
 def test_batch_chunking(batch_size):
-    latent_image = torch.randn(batch_size, 4, 64, 64).to(get_torch_device())
+    latent_image = torch.randn(batch_size, 4, 64, 64).to(CPU)
     target_shape = (4, 4, 64, 64)
 
     chunked = chunk_batch(latent_image, target_shape)
@@ -45,7 +42,7 @@ def test_batch_chunking(batch_size):
 
 @pytest.mark.parametrize("batch_size", [1, 2, 4, 5, 9])
 def test_merge_chunks(batch_size):
-    input_tensor = torch.randn(batch_size, 4, 64, 64).to(get_torch_device())
+    input_tensor = torch.randn(batch_size, 4, 64, 64).to(CPU)
     target_shape = (4, 4, 64, 64)
     chunked = chunk_batch(input_tensor, target_shape)
 
@@ -57,16 +54,16 @@ def test_merge_chunks(batch_size):
 
 @pytest.fixture
 def inputs():
-    x = torch.randn(1, 4, 64, 64).to(get_torch_device())
-    t = torch.randn([1]).to(get_torch_device())
-    c_crossattn = torch.randn(1, 77, 768).to(get_torch_device())
+    x = torch.randn(1, 4, 64, 64).to(CPU)
+    t = torch.randn([1]).to(CPU)
+    c_crossattn = torch.randn(1, 77, 768).to(CPU)
     control = {
         "output": [
-            torch.randn(1, 320, 64, 64).to(get_torch_device()),
-            torch.randn(1, 640, 32, 32).to(get_torch_device()),
+            torch.randn(1, 320, 64, 64).to(CPU),
+            torch.randn(1, 640, 32, 32).to(CPU),
         ],
     }
-    timestep_cond = torch.randn(1, 256).to(get_torch_device())
+    timestep_cond = torch.randn(1, 256).to(CPU)
 
     return CoreMLInputs(x, t, c_crossattn, control, timestep_cond=timestep_cond)
 
@@ -86,11 +83,11 @@ def inputs():
 def test_chunking_controlnet(b, target_size, num_chunks):
     cn = {
         "output": [
-            torch.randn(b, 320, 64, 64).to(get_torch_device()),
-            torch.randn(b, 640, 32, 32).to(get_torch_device()),
+            torch.randn(b, 320, 64, 64).to(CPU),
+            torch.randn(b, 640, 32, 32).to(CPU),
         ],
         "middle": [
-            torch.randn(b, 1280, 8, 8).to(get_torch_device()),
+            torch.randn(b, 1280, 8, 8).to(CPU),
         ],
     }
 

tests/unit/test_conversion_helpers.py

@@ -0,0 +1,183 @@
+from types import SimpleNamespace
+
+import torch
+
+from coreml_suite.conversion.attention import (
+    SplitEinsumAttnProcessor,
+    SplitEinsumV2AttnProcessor,
+    apply_attention_implementation,
+    split_einsum,
+    split_einsum_v2,
+)
+from coreml_suite.conversion.shapes import conv2d_output_shape
+from coreml_suite.conversion.unet import CoreMLUNetWrapper
+
+
+class RecordingUNet(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.call = None
+
+    def forward(
+        self,
+        sample,
+        timestep,
+        encoder_hidden_states,
+        timestep_cond=None,
+        added_cond_kwargs=None,
+        down_block_additional_residuals=None,
+        mid_block_additional_residual=None,
+        return_dict=True,
+        **kwargs,
+    ):
+        self.call = {
+            "sample": sample,
+            "timestep": timestep,
+            "encoder_hidden_states": encoder_hidden_states,
+            "timestep_cond": timestep_cond,
+            "added_cond_kwargs": added_cond_kwargs,
+            "down_block_additional_residuals": down_block_additional_residuals,
+            "mid_block_additional_residual": mid_block_additional_residual,
+            "return_dict": return_dict,
+        }
+        return (sample + 1,)
+
+
+def test_conv2d_output_shape_matches_torch_conv2d_contract():
+    conv = torch.nn.Conv2d(
+        4,
+        8,
+        kernel_size=(3, 5),
+        stride=(2, 3),
+        padding=(1, 2),
+        dilation=(1, 2),
+    )
+
+    assert conv2d_output_shape(17, 19, conv) == (9, 5)
+
+
+def test_unet_wrapper_passes_context_through_for_sd15():
+    unet = RecordingUNet()
+    wrapper = CoreMLUNetWrapper(unet, SimpleNamespace(name="SD15"))
+
+    sample = torch.randn(2, 4, 8, 8)
+    timestep = torch.randn(2)
+    context = torch.randn(2, 77, 768)
+
+    out = wrapper(sample, timestep, context)
+
+    assert torch.equal(out, sample + 1)
+    assert unet.call["encoder_hidden_states"] is context
+    assert unet.call["return_dict"] is False
+
+
+def test_unet_wrapper_routes_lcm_sdxl_and_controlnet_inputs():
+    unet = RecordingUNet()
+    wrapper = CoreMLUNetWrapper(unet, SimpleNamespace(name="LCM"))
+
+    sample = torch.randn(1, 4, 8, 8)
+    timestep = torch.randn(1)
+    context = torch.randn(1, 77, 768)
+    timestep_cond = torch.randn(1, 256)
+    down_residual = torch.randn(1, 320, 8, 8)
+    mid_residual = torch.randn(1, 1280, 1, 1)
+
+    wrapper(sample, timestep, context, timestep_cond, down_residual, mid_residual)
+
+    assert unet.call["timestep_cond"] is timestep_cond
+    assert len(unet.call["down_block_additional_residuals"]) == 1
+    assert unet.call["down_block_additional_residuals"][0] is down_residual
+    assert unet.call["mid_block_additional_residual"] is mid_residual
+
+
+def test_unet_wrapper_routes_sdxl_added_conditioning():
+    unet = RecordingUNet()
+    wrapper = CoreMLUNetWrapper(unet, SimpleNamespace(name="SDXL"))
+
+    sample = torch.randn(1, 4, 8, 8)
+    timestep = torch.randn(1)
+    context = torch.randn(1, 77, 2048)
+    time_ids = torch.randn(1, 6)
+    text_embeds = torch.randn(1, 1280)
+
+    wrapper(sample, timestep, context, time_ids, text_embeds)
+
+    assert unet.call["added_cond_kwargs"]["time_ids"] is time_ids
+    assert unet.call["added_cond_kwargs"]["text_embeds"] is text_embeds
+
+
+def test_split_einsum_matches_original_attention_math():
+    torch.manual_seed(0)
+    batch = 2
+    heads = 3
+    dim_head = 4
+    sequence = 16
+    channels = heads * dim_head
+    q = torch.randn(batch, channels, 1, sequence)
+    k = torch.randn(batch, channels, 1, sequence)
+    v = torch.randn(batch, channels, 1, sequence)
+
+    expected = _original_attention(q, k, v, None, heads, dim_head)
+
+    # split-einsum reorders the float32 reductions vs the reference, so equality
+    # only holds up to rounding; the drift exceeds allclose's default atol on
+    # some BLAS backends (e.g. Linux x86 CI).
+    assert torch.allclose(split_einsum(q, k, v, None, heads, dim_head), expected, atol=1e-6)
+    assert torch.allclose(split_einsum_v2(q, k, v, None, heads, dim_head), expected, atol=1e-6)
+
+
+def test_split_einsum_v2_chunked_path_matches_original_attention_math():
+    torch.manual_seed(0)
+    batch = 1
+    heads = 2
+    dim_head = 2
+    sequence = 512
+    channels = heads * dim_head
+    q = torch.randn(batch, channels, 1, sequence)
+    k = torch.randn(batch, channels, 1, sequence)
+    v = torch.randn(batch, channels, 1, sequence)
+
+    expected = _original_attention(q, k, v, None, heads, dim_head)
+
+    assert torch.allclose(
+        split_einsum_v2(q, k, v, None, heads, dim_head),
+        expected,
+        atol=1e-6,
+    )
+
+
+def test_apply_attention_implementation_sets_split_processors():
+    unet = RecordingProcessorUNet()
+
+    assert apply_attention_implementation(unet, "ORIGINAL") is unet
+    assert unet.processor is None
+
+    apply_attention_implementation(unet, "SPLIT_EINSUM")
+    assert isinstance(unet.processor, SplitEinsumAttnProcessor)
+
+    apply_attention_implementation(unet, "SPLIT_EINSUM_V2")
+    assert isinstance(unet.processor, SplitEinsumV2AttnProcessor)
+
+
+class RecordingProcessorUNet:
+    def __init__(self):
+        self.processor = None
+
+    def set_attn_processor(self, processor):
+        self.processor = processor
+
+
+def _original_attention(q, k, v, mask, heads, dim_head):
+    batch = q.size(0)
+    mh_q = q.view(batch, heads, dim_head, -1)
+    mh_k = k.view(batch, heads, dim_head, -1)
+    mh_v = v.view(batch, heads, dim_head, -1)
+
+    weights = torch.einsum("bhcq,bhck->bhqk", mh_q, mh_k)
+    weights = weights * (dim_head**-0.5)
+    if mask is not None:
+        weights = weights + mask
+    weights = weights.softmax(dim=3)
+
+    attn = torch.einsum("bhqk,bhck->bhcq", weights, mh_v)
+    return attn.contiguous().view(batch, heads * dim_head, 1, -1)

tests/unit/test_tier0_purity.py

@@ -0,0 +1,43 @@
+"""Gate: prove the Tier-0 lane is framework-free.
+
+In a pure `pytest -m unit` run, none of the banned runtime modules
+(comfy, coremltools, python_coreml_stable_diffusion, folder_paths,
+nodes, comfy_extras, diffusers, diffusionkit) may be in sys.modules
+after collection. If they are, a tests/unit/ file is transitively
+pulling them in and the Tier-0 promise — "runs on Linux with no Mac
+stack" — is broken.
+
+When other tiers are also collected, framework modules may be imported
+deliberately (e.g. smoke pulls in coremltools), so the check is skipped
+unless the run is purely `-m unit` — Tier-0 purity is only meaningful
+when nothing else is loaded.
+"""
+import sys
+
+import pytest
+
+BANNED_ROOTS = {
+    "comfy",
+    "comfy_extras",
+    "coremltools",
+    "python_coreml_stable_diffusion",
+    "folder_paths",
+    "nodes",
+    "diffusers",
+    "diffusionkit",
+}
+
+
+def test_no_framework_modules_loaded_by_unit_tier(request):
+    markexpr = request.config.option.markexpr
+    if markexpr != "unit":
+        pytest.skip(
+            "purity gate only meaningful in a pure `-m unit` run "
+            f"(got markexpr={markexpr!r}); other tiers are expected to "
+            "import comfy/coremltools."
+        )
+    loaded = {name for name in sys.modules if name.split(".")[0] in BANNED_ROOTS}
+    assert not loaded, (
+        f"Tier-0 leakage: these framework modules are in sys.modules after "
+        f"collecting tests/unit/: {sorted(loaded)}. Pure-core promise broken."
+    )